iobroker.assistant 1.4.0 → 1.4.2

package/build/lib/voice/fallback.js

@@ -1,12 +1,24 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.FallbackTts = exports.FallbackStt = void 0;
-/** Run `attempt` over the engines in order; the last error survives if all of them fail. */
-async function overEngines(engines, what, log, attempt) {
+/**
+ * Run `attempt` over the engines in order; the last error survives if all of them fail. A result that
+ * `useful` rejects also moves on to the next engine; if none does better, the first such result is returned.
+ */
+async function overEngines(engines, what, log, attempt, useful) {
     let last;
+    let unused;
     for (const [index, entry] of engines.entries()) {
         try {
             const result = await attempt(entry.engine);
+            const next = engines[index + 1];
+            if (useful && !useful(result)) {
+                unused ??= { result };
+                if (next) {
+                    log.info(`${what}: ${entry.name} returned nothing — trying ${next.name}.`);
+                }
+                continue;
+            }
             if (index > 0) {
                 log.info(`${what}: answered by the fallback engine (${entry.name}).`);
             }
@@ -18,6 +30,9 @@ async function overEngines(engines, what, log, attempt) {
             log[next ? 'warn' : 'error'](`${what}: ${entry.name} failed (${last.message})${next ? ` — trying ${next.name}` : ''}.`);
         }
     }
+    if (unused) {
+        return unused.result;
+    }
     throw last ?? new Error(`${what}: no engine configured`);
 }
 /** Speech-to-text with one fallback engine. */
@@ -29,7 +44,7 @@ class FallbackStt {
         this.log = log;
     }
     transcribe(pcm, sampleRate, lang, hints) {
-        return overEngines(this.engines, 'Speech-to-text', this.log, e => e.transcribe(pcm, sampleRate, lang, hints));
+        return overEngines(this.engines, 'Speech-to-text', this.log, e => e.transcribe(pcm, sampleRate, lang, hints), text => !!text.trim());
     }
     /** Prepare every engine: the fallback is useless if its model is only downloaded once it is needed. */
     async prepare(lang) {

package/build/lib/voice/fallback.js.map

@@ -1 +1 @@
-{"version":3,"file":"fallback.js","sourceRoot":"","sources":["../../../src/lib/voice/fallback.ts"],"names":[],"mappings":";;;AA0BA,4FAA4F;AAC5F,KAAK,UAAU,WAAW,CACtB,OAAyB,EACzB,IAAY,EACZ,GAAgB,EAChB,OAAkC;IAElC,IAAI,IAAuB,CAAC;IAC5B,KAAK,MAAM,CAAC,KAAK,EAAE,KAAK,CAAC,IAAI,OAAO,CAAC,OAAO,EAAE,EAAE,CAAC;QAC7C,IAAI,CAAC;YACD,MAAM,MAAM,GAAG,MAAM,OAAO,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;YAC3C,IAAI,KAAK,GAAG,CAAC,EAAE,CAAC;gBACZ,GAAG,CAAC,IAAI,CAAC,GAAG,IAAI,sCAAsC,KAAK,CAAC,IAAI,IAAI,CAAC,CAAC;YAC1E,CAAC;YACD,OAAO,MAAM,CAAC;QAClB,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACT,IAAI,GAAG,CAAU,CAAC;YAClB,MAAM,IAAI,GAAG,OAAO,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;YAChC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,CACxB,GAAG,IAAI,KAAK,KAAK,CAAC,IAAI,YAAY,IAAI,CAAC,OAAO,IAAI,IAAI,CAAC,CAAC,CAAC,aAAa,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,GAAG,CAC5F,CAAC;QACN,CAAC;IACL,CAAC;IACD,MAAM,IAAI,IAAI,IAAI,KAAK,CAAC,GAAG,IAAI,wBAAwB,CAAC,CAAC;AAC7D,CAAC;AAED,+CAA+C;AAC/C,MAAa,WAAW;IAEC;IACA;IAFrB,YACqB,OAAiC,EACjC,GAAgB;QADhB,YAAO,GAAP,OAAO,CAA0B;QACjC,QAAG,GAAH,GAAG,CAAa;IAClC,CAAC;IAEJ,UAAU,CAAC,GAAW,EAAE,UAAkB,EAAE,IAAY,EAAE,KAAgB;QACtE,OAAO,WAAW,CAAC,IAAI,CAAC,OAAO,EAAE,gBAAgB,EAAE,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,UAAU,CAAC,GAAG,EAAE,UAAU,EAAE,IAAI,EAAE,KAAK,CAAC,CAAC,CAAC;IAClH,CAAC;IAED,uGAAuG;IACvG,KAAK,CAAC,OAAO,CAAC,IAAY;QACtB,KAAK,MAAM,KAAK,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YAC/B,IAAI,CAAC;gBACD,MAAM,KAAK,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC,IAAI,CAAC,CAAC;YACvC,CAAC;YAAC,OAAO,CAAC,EAAE,CAAC;gBACT,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,6BAA6B,KAAK,CAAC,IAAI,YAAa,CAAW,CAAC,OAAO,IAAI,CAAC,CAAC;YAC/F,CAAC;QACL,CAAC;IACL,CAAC;CACJ;AApBD,kCAoBC;AAED,+CAA+C;AAC/C,MAAa,WAAW;IAEC;IACA;IAFrB,YACqB,OAAiC,EACjC,GAAgB;QADhB,YAAO,GAAP,OAAO,CAA0B;QACjC,QAAG,GAAH,GAAG,CAAa;IAClC,CAAC;IAEJ,UAAU,CAAC,IAAY,EAAE,IAAY;QACjC,OAAO,WAAW,CAAC,IAAI,CAAC,OAAO,EAAE,gBAAgB,EAAE,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,UAAU,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC;IAChG,CAAC;IAED;;;OAGG;IACH,cAAc,CAAC,IAAY,EAAE,IAAY;QACrC,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,cAAc,CAAC,CAAC;QACpE,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,CAAC;YAClB,OAAO,IAAI,CAAC,UAAU,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;QACvC,CAAC;QACD,OAAO,WAAW,CAAC,OAAO,EAAE,uBAAuB,EAAE,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,cAAe,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC;IACvG,CAAC;IAED,KAAK,CAAC,OAAO,CAAC,IAAY;QACtB,KAAK,MAAM,KAAK,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YAC/B,IAAI,CAAC;gBACD,MAAM,KAAK,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC,IAAI,CAAC,CAAC;YACvC,CAAC;YAAC,OAAO,CAAC,EAAE,CAAC;gBACT,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,6BAA6B,KAAK,CAAC,IAAI,YAAa,CAAW,CAAC,OAAO,IAAI,CAAC,CAAC;YAC/F,CAAC;QACL,CAAC;IACL,CAAC;IAED,8FAA8F;IAC9F,IAAI,CAAC,OAAiB,EAAE,IAAY;QAChC,OAAO,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,MAAM,CAAC,IAAI,EAAE,CAAC,OAAO,EAAE,IAAI,CAAC,IAAI,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC;IAC/E,CAAC;CACJ;AApCD,kCAoCC","sourcesContent":["/**\n * A second speech engine behind the first one, so an outage does not take the voice away.\n *\n * The usual setup is a cloud engine with a local one behind it: when OpenAI/Azure/AWS is unreachable, the\n * API key expired or the quota ran out, Vosk and Piper keep the house answering — worse sounding, but\n * answering. That is the whole value: without it, the assistant goes mute exactly when the internet does.\n *\n * The idea comes from the Python original (`C:\\iot\\Hannah`, `core/hannah/stt.py`, which chains\n * aws → azure → remote → local); here it is one explicit fallback rather than an open chain, because a\n * household configures at most two speech providers and a longer chain would mostly be a way to wait\n * several timeouts in a row.\n *\n * Any failure of the primary moves on to the fallback: the alternative would be classifying provider\n * errors into \"retryable\" and \"not\", which is guesswork across three SDKs — and a request the primary\n * cannot answer is one the user still wants answered.\n */\nimport type { SttEngine } from './stt';\nimport type { TtsEngine, TtsResult } from './tts';\nimport type { VoiceLogger } from './download';\n\n/** One named engine in the chain. */\nexport interface NamedEngine<T> {\n    name: string;\n    engine: T;\n}\n\n/** Run `attempt` over the engines in order; the last error survives if all of them fail. */\nasync function overEngines<T, R>(\n    engines: NamedEngine<T>[],\n    what: string,\n    log: VoiceLogger,\n    attempt: (engine: T) => Promise<R>,\n): Promise<R> {\n    let last: Error | undefined;\n    for (const [index, entry] of engines.entries()) {\n        try {\n            const result = await attempt(entry.engine);\n            if (index > 0) {\n                log.info(`${what}: answered by the fallback engine (${entry.name}).`);\n            }\n            return result;\n        } catch (e) {\n            last = e as Error;\n            const next = engines[index + 1];\n            log[next ? 'warn' : 'error'](\n                `${what}: ${entry.name} failed (${last.message})${next ? ` — trying ${next.name}` : ''}.`,\n            );\n        }\n    }\n    throw last ?? new Error(`${what}: no engine configured`);\n}\n\n/** Speech-to-text with one fallback engine. */\nexport class FallbackStt implements SttEngine {\n    constructor(\n        private readonly engines: NamedEngine<SttEngine>[],\n        private readonly log: VoiceLogger,\n    ) {}\n\n    transcribe(pcm: Buffer, sampleRate: number, lang: string, hints?: string[]): Promise<string> {\n        return overEngines(this.engines, 'Speech-to-text', this.log, e => e.transcribe(pcm, sampleRate, lang, hints));\n    }\n\n    /** Prepare every engine: the fallback is useless if its model is only downloaded once it is needed. */\n    async prepare(lang: string): Promise<void> {\n        for (const entry of this.engines) {\n            try {\n                await entry.engine.prepare?.(lang);\n            } catch (e) {\n                this.log.warn(`Speech-to-text: preparing ${entry.name} failed (${(e as Error).message}).`);\n            }\n        }\n    }\n}\n\n/** Text-to-speech with one fallback engine. */\nexport class FallbackTts implements TtsEngine {\n    constructor(\n        private readonly engines: NamedEngine<TtsEngine>[],\n        private readonly log: VoiceLogger,\n    ) {}\n\n    synthesize(text: string, lang: string): Promise<TtsResult> {\n        return overEngines(this.engines, 'Text-to-speech', this.log, e => e.synthesize(text, lang));\n    }\n\n    /**\n     * SSML only goes to engines that can speak it; if none can, the primary's plain-text path handles it\n     * (its cache wrapper strips the markup).\n     */\n    synthesizeSsml(ssml: string, lang: string): Promise<TtsResult> {\n        const capable = this.engines.filter(e => !!e.engine.synthesizeSsml);\n        if (!capable.length) {\n            return this.synthesize(ssml, lang);\n        }\n        return overEngines(capable, 'Text-to-speech (SSML)', this.log, e => e.synthesizeSsml!(ssml, lang));\n    }\n\n    async prepare(lang: string): Promise<void> {\n        for (const entry of this.engines) {\n            try {\n                await entry.engine.prepare?.(lang);\n            } catch (e) {\n                this.log.warn(`Text-to-speech: preparing ${entry.name} failed (${(e as Error).message}).`);\n            }\n        }\n    }\n\n    /** Warm only the primary's cache: the fallback is for an outage, not for everyday phrases. */\n    warm(phrases: string[], lang: string): Promise<number> {\n        return this.engines[0]?.engine.warm?.(phrases, lang) ?? Promise.resolve(0);\n    }\n}\n"]}
+{"version":3,"file":"fallback.js","sourceRoot":"","sources":["../../../src/lib/voice/fallback.ts"],"names":[],"mappings":";;;AA8BA;;;GAGG;AACH,KAAK,UAAU,WAAW,CACtB,OAAyB,EACzB,IAAY,EACZ,GAAgB,EAChB,OAAkC,EAClC,MAA+B;IAE/B,IAAI,IAAuB,CAAC;IAC5B,IAAI,MAAiC,CAAC;IACtC,KAAK,MAAM,CAAC,KAAK,EAAE,KAAK,CAAC,IAAI,OAAO,CAAC,OAAO,EAAE,EAAE,CAAC;QAC7C,IAAI,CAAC;YACD,MAAM,MAAM,GAAG,MAAM,OAAO,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;YAC3C,MAAM,IAAI,GAAG,OAAO,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;YAChC,IAAI,MAAM,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,CAAC;gBAC5B,MAAM,KAAK,EAAE,MAAM,EAAE,CAAC;gBACtB,IAAI,IAAI,EAAE,CAAC;oBACP,GAAG,CAAC,IAAI,CAAC,GAAG,IAAI,KAAK,KAAK,CAAC,IAAI,8BAA8B,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC;gBAC/E,CAAC;gBACD,SAAS;YACb,CAAC;YACD,IAAI,KAAK,GAAG,CAAC,EAAE,CAAC;gBACZ,GAAG,CAAC,IAAI,CAAC,GAAG,IAAI,sCAAsC,KAAK,CAAC,IAAI,IAAI,CAAC,CAAC;YAC1E,CAAC;YACD,OAAO,MAAM,CAAC;QAClB,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACT,IAAI,GAAG,CAAU,CAAC;YAClB,MAAM,IAAI,GAAG,OAAO,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;YAChC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,CACxB,GAAG,IAAI,KAAK,KAAK,CAAC,IAAI,YAAY,IAAI,CAAC,OAAO,IAAI,IAAI,CAAC,CAAC,CAAC,aAAa,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,GAAG,CAC5F,CAAC;QACN,CAAC;IACL,CAAC;IACD,IAAI,MAAM,EAAE,CAAC;QACT,OAAO,MAAM,CAAC,MAAM,CAAC;IACzB,CAAC;IACD,MAAM,IAAI,IAAI,IAAI,KAAK,CAAC,GAAG,IAAI,wBAAwB,CAAC,CAAC;AAC7D,CAAC;AAED,+CAA+C;AAC/C,MAAa,WAAW;IAEC;IACA;IAFrB,YACqB,OAAiC,EACjC,GAAgB;QADhB,YAAO,GAAP,OAAO,CAA0B;QACjC,QAAG,GAAH,GAAG,CAAa;IAClC,CAAC;IAEJ,UAAU,CAAC,GAAW,EAAE,UAAkB,EAAE,IAAY,EAAE,KAAgB;QACtE,OAAO,WAAW,CACd,IAAI,CAAC,OAAO,EACZ,gBAAgB,EAChB,IAAI,CAAC,GAAG,EACR,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,UAAU,CAAC,GAAG,EAAE,UAAU,EAAE,IAAI,EAAE,KAAK,CAAC,EAC/C,IAAI,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,EAAE,CACxB,CAAC;IACN,CAAC;IAED,uGAAuG;IACvG,KAAK,CAAC,OAAO,CAAC,IAAY;QACtB,KAAK,MAAM,KAAK,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YAC/B,IAAI,CAAC;gBACD,MAAM,KAAK,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC,IAAI,CAAC,CAAC;YACvC,CAAC;YAAC,OAAO,CAAC,EAAE,CAAC;gBACT,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,6BAA6B,KAAK,CAAC,IAAI,YAAa,CAAW,CAAC,OAAO,IAAI,CAAC,CAAC;YAC/F,CAAC;QACL,CAAC;IACL,CAAC;CACJ;AA1BD,kCA0BC;AAED,+CAA+C;AAC/C,MAAa,WAAW;IAEC;IACA;IAFrB,YACqB,OAAiC,EACjC,GAAgB;QADhB,YAAO,GAAP,OAAO,CAA0B;QACjC,QAAG,GAAH,GAAG,CAAa;IAClC,CAAC;IAEJ,UAAU,CAAC,IAAY,EAAE,IAAY;QACjC,OAAO,WAAW,CAAC,IAAI,CAAC,OAAO,EAAE,gBAAgB,EAAE,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,UAAU,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC;IAChG,CAAC;IAED;;;OAGG;IACH,cAAc,CAAC,IAAY,EAAE,IAAY;QACrC,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,cAAc,CAAC,CAAC;QACpE,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,CAAC;YAClB,OAAO,IAAI,CAAC,UAAU,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;QACvC,CAAC;QACD,OAAO,WAAW,CAAC,OAAO,EAAE,uBAAuB,EAAE,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,cAAe,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC;IACvG,CAAC;IAED,KAAK,CAAC,OAAO,CAAC,IAAY;QACtB,KAAK,MAAM,KAAK,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YAC/B,IAAI,CAAC;gBACD,MAAM,KAAK,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC,IAAI,CAAC,CAAC;YACvC,CAAC;YAAC,OAAO,CAAC,EAAE,CAAC;gBACT,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,6BAA6B,KAAK,CAAC,IAAI,YAAa,CAAW,CAAC,OAAO,IAAI,CAAC,CAAC;YAC/F,CAAC;QACL,CAAC;IACL,CAAC;IAED,8FAA8F;IAC9F,IAAI,CAAC,OAAiB,EAAE,IAAY;QAChC,OAAO,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,MAAM,CAAC,IAAI,EAAE,CAAC,OAAO,EAAE,IAAI,CAAC,IAAI,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC;IAC/E,CAAC;CACJ;AApCD,kCAoCC","sourcesContent":["/**\n * A second speech engine behind the first one, so an outage does not take the voice away.\n *\n * The usual setup is a cloud engine with a local one behind it: when OpenAI/Azure/AWS is unreachable, the\n * API key expired or the quota ran out, Vosk and Piper keep the house answering — worse sounding, but\n * answering. That is the whole value: without it, the assistant goes mute exactly when the internet does.\n *\n * The idea comes from the Python original (`C:\\iot\\Hannah`, `core/hannah/stt.py`, which chains\n * aws → azure → remote → local); here it is one explicit fallback rather than an open chain, because a\n * household configures at most two speech providers and a longer chain would mostly be a way to wait\n * several timeouts in a row.\n *\n * Any failure of the primary moves on to the fallback: the alternative would be classifying provider\n * errors into \"retryable\" and \"not\", which is guesswork across three SDKs — and a request the primary\n * cannot answer is one the user still wants answered.\n *\n * Speech-to-text also moves on when the primary returns no text at all. A satellite only sends audio it took\n * for speech, and models differ in what they make of it: Nemotron returned nothing for a quiet\n * \"Мессенджер. Да, да.\" that Whisper recognised. The price is a second request for a false wake.\n */\nimport type { SttEngine } from './stt';\nimport type { TtsEngine, TtsResult } from './tts';\nimport type { VoiceLogger } from './download';\n\n/** One named engine in the chain. */\nexport interface NamedEngine<T> {\n    name: string;\n    engine: T;\n}\n\n/**\n * Run `attempt` over the engines in order; the last error survives if all of them fail. A result that\n * `useful` rejects also moves on to the next engine; if none does better, the first such result is returned.\n */\nasync function overEngines<T, R>(\n    engines: NamedEngine<T>[],\n    what: string,\n    log: VoiceLogger,\n    attempt: (engine: T) => Promise<R>,\n    useful?: (result: R) => boolean,\n): Promise<R> {\n    let last: Error | undefined;\n    let unused: { result: R } | undefined;\n    for (const [index, entry] of engines.entries()) {\n        try {\n            const result = await attempt(entry.engine);\n            const next = engines[index + 1];\n            if (useful && !useful(result)) {\n                unused ??= { result };\n                if (next) {\n                    log.info(`${what}: ${entry.name} returned nothing — trying ${next.name}.`);\n                }\n                continue;\n            }\n            if (index > 0) {\n                log.info(`${what}: answered by the fallback engine (${entry.name}).`);\n            }\n            return result;\n        } catch (e) {\n            last = e as Error;\n            const next = engines[index + 1];\n            log[next ? 'warn' : 'error'](\n                `${what}: ${entry.name} failed (${last.message})${next ? ` — trying ${next.name}` : ''}.`,\n            );\n        }\n    }\n    if (unused) {\n        return unused.result;\n    }\n    throw last ?? new Error(`${what}: no engine configured`);\n}\n\n/** Speech-to-text with one fallback engine. */\nexport class FallbackStt implements SttEngine {\n    constructor(\n        private readonly engines: NamedEngine<SttEngine>[],\n        private readonly log: VoiceLogger,\n    ) {}\n\n    transcribe(pcm: Buffer, sampleRate: number, lang: string, hints?: string[]): Promise<string> {\n        return overEngines(\n            this.engines,\n            'Speech-to-text',\n            this.log,\n            e => e.transcribe(pcm, sampleRate, lang, hints),\n            text => !!text.trim(),\n        );\n    }\n\n    /** Prepare every engine: the fallback is useless if its model is only downloaded once it is needed. */\n    async prepare(lang: string): Promise<void> {\n        for (const entry of this.engines) {\n            try {\n                await entry.engine.prepare?.(lang);\n            } catch (e) {\n                this.log.warn(`Speech-to-text: preparing ${entry.name} failed (${(e as Error).message}).`);\n            }\n        }\n    }\n}\n\n/** Text-to-speech with one fallback engine. */\nexport class FallbackTts implements TtsEngine {\n    constructor(\n        private readonly engines: NamedEngine<TtsEngine>[],\n        private readonly log: VoiceLogger,\n    ) {}\n\n    synthesize(text: string, lang: string): Promise<TtsResult> {\n        return overEngines(this.engines, 'Text-to-speech', this.log, e => e.synthesize(text, lang));\n    }\n\n    /**\n     * SSML only goes to engines that can speak it; if none can, the primary's plain-text path handles it\n     * (its cache wrapper strips the markup).\n     */\n    synthesizeSsml(ssml: string, lang: string): Promise<TtsResult> {\n        const capable = this.engines.filter(e => !!e.engine.synthesizeSsml);\n        if (!capable.length) {\n            return this.synthesize(ssml, lang);\n        }\n        return overEngines(capable, 'Text-to-speech (SSML)', this.log, e => e.synthesizeSsml!(ssml, lang));\n    }\n\n    async prepare(lang: string): Promise<void> {\n        for (const entry of this.engines) {\n            try {\n                await entry.engine.prepare?.(lang);\n            } catch (e) {\n                this.log.warn(`Text-to-speech: preparing ${entry.name} failed (${(e as Error).message}).`);\n            }\n        }\n    }\n\n    /** Warm only the primary's cache: the fallback is for an outage, not for everyday phrases. */\n    warm(phrases: string[], lang: string): Promise<number> {\n        return this.engines[0]?.engine.warm?.(phrases, lang) ?? Promise.resolve(0);\n    }\n}\n"]}

package/build/lib/voice/speechAudio.js

@@ -0,0 +1,97 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.MAX_GAIN = exports.TARGET_PEAK = exports.TAIL_SHARE = exports.TAIL_MS = exports.PRE_ROLL_MS = void 0;
+exports.prepareForStt = prepareForStt;
+/**
+ * Audio for speech recognition: the recording cut to the speech and brought to a usable level.
+ *
+ * ESPHome satellites stream from the wake word on, so a recording starts with the device's own chirp or beep
+ * and the silence before anybody speaks, and ends with the silence that told us the speaker stopped. Measured
+ * on a ThirdReality speaker (2026-11-03): the answer to a follow-up question — a 250 ms beep, a second of
+ * digital silence, then "Мессенджер. Да, да." at RMS 150–500 — came back empty from Nemotron, and Whisper
+ * made "Продолжение следует... Мессенджет. и я" of it (the subtitle line Whisper invents for silence). Cut to
+ * the speech and amplified, Whisper heard it right.
+ *
+ * Only what goes to the recognition changes; the diagnosis recordings keep the audio as it arrived.
+ */
+const vad_1 = require("./vad");
+/** Audio kept before the frame where speech was detected: the soft start of a word lies below the gate. */
+exports.PRE_ROLL_MS = 400;
+/** Audio kept after the last sound: the fading end of a word. */
+exports.TAIL_MS = 400;
+/**
+ * A frame in the tail counts as sound when it reaches this share of the silence threshold — low, so that a
+ * quiet last syllable stays in. In a noisy room nothing is quiet enough, and the tail is simply kept.
+ */
+exports.TAIL_SHARE = 0.4;
+/** Peak sample quiet speech is raised to — about half of full scale, headroom for the loudest syllable. */
+exports.TARGET_PEAK = 16000;
+/** Never amplify more than this: beyond it only the noise grows. */
+exports.MAX_GAIN = 10;
+/** Gains below this are not worth touching the audio for. */
+const MIN_GAIN = 1.25;
+/** Never cut a recording shorter than this — a short "да" plus its margins. */
+const MIN_KEEP_MS = 600;
+/** Resolution of the tail search. */
+const FRAME_MS = 30;
+/**
+ * Cut `pcm` (mono 16-bit) to the speech and amplify it when it is quiet.
+ *
+ * @param pcm the whole recording
+ * @param sampleRate its sample rate
+ * @param speech what the detector saw
+ * @param speech.startMs where speech started, in ms (absent = keep the beginning)
+ * @param speech.endLevel the silence threshold; the tail keeps everything above 40 % of it
+ */
+function prepareForStt(pcm, sampleRate, speech) {
+    const samples = Math.floor(pcm.length / 2);
+    const perMs = sampleRate / 1000;
+    const totalMs = Math.floor(samples / perMs);
+    let fromMs = speech.startMs !== undefined && speech.startMs >= 0 ? Math.max(0, speech.startMs - exports.PRE_ROLL_MS) : 0;
+    fromMs = Math.min(fromMs, Math.max(0, totalMs - MIN_KEEP_MS));
+    // The tail: the last frame with any sound in it, plus a margin.
+    let toMs = totalMs;
+    const quiet = speech.endLevel * exports.TAIL_SHARE;
+    if (quiet > 0) {
+        const frame = Math.max(1, Math.round(FRAME_MS * perMs));
+        const first = Math.round(fromMs * perMs);
+        let lastSound = -1;
+        for (let start = samples - frame; start >= first; start -= frame) {
+            if ((0, vad_1.rms)(pcm.subarray(start * 2, (start + frame) * 2)) >= quiet) {
+                lastSound = start + frame;
+                break;
+            }
+        }
+        if (lastSound >= 0) {
+            toMs = Math.min(totalMs, Math.ceil(lastSound / perMs) + exports.TAIL_MS);
+        }
+    }
+    if (toMs - fromMs < MIN_KEEP_MS) {
+        toMs = Math.min(totalMs, fromMs + MIN_KEEP_MS);
+    }
+    const from = Math.round(fromMs * perMs);
+    const to = Math.min(samples, Math.round(toMs * perMs));
+    let peak = 0;
+    for (let i = from; i < to; i++) {
+        const value = Math.abs(pcm.readInt16LE(i * 2));
+        if (value > peak) {
+            peak = value;
+        }
+    }
+    let gain = peak > 0 ? Math.min(exports.MAX_GAIN, exports.TARGET_PEAK / peak) : 1;
+    if (gain < MIN_GAIN) {
+        gain = 1;
+    }
+    const out = Buffer.alloc((to - from) * 2);
+    if (gain === 1) {
+        pcm.copy(out, 0, from * 2, to * 2);
+    }
+    else {
+        for (let i = from; i < to; i++) {
+            const value = Math.round(pcm.readInt16LE(i * 2) * gain);
+            out.writeInt16LE(Math.max(-32768, Math.min(32767, value)), (i - from) * 2);
+        }
+    }
+    return { pcm: out, fromMs, toMs, gain: Math.round(gain * 100) / 100 };
+}
+//# sourceMappingURL=speechAudio.js.map

package/build/lib/voice/speechAudio.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"speechAudio.js","sourceRoot":"","sources":["../../../src/lib/voice/speechAudio.ts"],"names":[],"mappings":";;;AAqDA,sCAyDC;AA9GD;;;;;;;;;;;GAWG;AACH,+BAAwC;AAExC,2GAA2G;AAC9F,QAAA,WAAW,GAAG,GAAG,CAAC;AAC/B,iEAAiE;AACpD,QAAA,OAAO,GAAG,GAAG,CAAC;AAC3B;;;GAGG;AACU,QAAA,UAAU,GAAG,GAAG,CAAC;AAC9B,2GAA2G;AAC9F,QAAA,WAAW,GAAG,KAAK,CAAC;AACjC,oEAAoE;AACvD,QAAA,QAAQ,GAAG,EAAE,CAAC;AAC3B,6DAA6D;AAC7D,MAAM,QAAQ,GAAG,IAAI,CAAC;AACtB,+EAA+E;AAC/E,MAAM,WAAW,GAAG,GAAG,CAAC;AACxB,qCAAqC;AACrC,MAAM,QAAQ,GAAG,EAAE,CAAC;AAYpB;;;;;;;;GAQG;AACH,SAAgB,aAAa,CACzB,GAAU,EACV,UAAkB,EAClB,MAA8C;IAE9C,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;IAC3C,MAAM,KAAK,GAAG,UAAU,GAAG,IAAI,CAAC;IAChC,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,OAAO,GAAG,KAAK,CAAC,CAAC;IAE5C,IAAI,MAAM,GAAG,MAAM,CAAC,OAAO,KAAK,SAAS,IAAI,MAAM,CAAC,OAAO,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,MAAM,CAAC,OAAO,GAAG,mBAAW,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACjH,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,GAAG,WAAW,CAAC,CAAC,CAAC;IAE9D,gEAAgE;IAChE,IAAI,IAAI,GAAG,OAAO,CAAC;IACnB,MAAM,KAAK,GAAG,MAAM,CAAC,QAAQ,GAAG,kBAAU,CAAC;IAC3C,IAAI,KAAK,GAAG,CAAC,EAAE,CAAC;QACZ,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,QAAQ,GAAG,KAAK,CAAC,CAAC,CAAC;QACxD,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,GAAG,KAAK,CAAC,CAAC;QACzC,IAAI,SAAS,GAAG,CAAC,CAAC,CAAC;QACnB,KAAK,IAAI,KAAK,GAAG,OAAO,GAAG,KAAK,EAAE,KAAK,IAAI,KAAK,EAAE,KAAK,IAAI,KAAK,EAAE,CAAC;YAC/D,IAAI,IAAA,SAAG,EAAC,GAAG,CAAC,QAAQ,CAAC,KAAK,GAAG,CAAC,EAAE,CAAC,KAAK,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,KAAK,EAAE,CAAC;gBAC7D,SAAS,GAAG,KAAK,GAAG,KAAK,CAAC;gBAC1B,MAAM;YACV,CAAC;QACL,CAAC;QACD,IAAI,SAAS,IAAI,CAAC,EAAE,CAAC;YACjB,IAAI,GAAG,IAAI,CAAC,GAAG,CAAC,OAAO,EAAE,IAAI,CAAC,IAAI,CAAC,SAAS,GAAG,KAAK,CAAC,GAAG,eAAO,CAAC,CAAC;QACrE,CAAC;IACL,CAAC;IACD,IAAI,IAAI,GAAG,MAAM,GAAG,WAAW,EAAE,CAAC;QAC9B,IAAI,GAAG,IAAI,CAAC,GAAG,CAAC,OAAO,EAAE,MAAM,GAAG,WAAW,CAAC,CAAC;IACnD,CAAC;IAED,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,GAAG,KAAK,CAAC,CAAC;IACxC,MAAM,EAAE,GAAG,IAAI,CAAC,GAAG,CAAC,OAAO,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,GAAG,KAAK,CAAC,CAAC,CAAC;IACvD,IAAI,IAAI,GAAG,CAAC,CAAC;IACb,KAAK,IAAI,CAAC,GAAG,IAAI,EAAE,CAAC,GAAG,EAAE,EAAE,CAAC,EAAE,EAAE,CAAC;QAC7B,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,WAAW,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;QAC/C,IAAI,KAAK,GAAG,IAAI,EAAE,CAAC;YACf,IAAI,GAAG,KAAK,CAAC;QACjB,CAAC;IACL,CAAC;IACD,IAAI,IAAI,GAAG,IAAI,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,gBAAQ,EAAE,mBAAW,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACjE,IAAI,IAAI,GAAG,QAAQ,EAAE,CAAC;QAClB,IAAI,GAAG,CAAC,CAAC;IACb,CAAC;IAED,MAAM,GAAG,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC;IAC1C,IAAI,IAAI,KAAK,CAAC,EAAE,CAAC;QACb,GAAG,CAAC,IAAI,CAAC,GAAG,EAAE,CAAC,EAAE,IAAI,GAAG,CAAC,EAAE,EAAE,GAAG,CAAC,CAAC,CAAC;IACvC,CAAC;SAAM,CAAC;QACJ,KAAK,IAAI,CAAC,GAAG,IAAI,EAAE,CAAC,GAAG,EAAE,EAAE,CAAC,EAAE,EAAE,CAAC;YAC7B,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,WAAW,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC;YACxD,GAAG,CAAC,YAAY,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC;QAC/E,CAAC;IACL,CAAC;IACD,OAAO,EAAE,GAAG,EAAE,GAAG,EAAE,MAAM,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,GAAG,GAAG,CAAC,GAAG,GAAG,EAAE,CAAC;AAC1E,CAAC","sourcesContent":["/**\n * Audio for speech recognition: the recording cut to the speech and brought to a usable level.\n *\n * ESPHome satellites stream from the wake word on, so a recording starts with the device's own chirp or beep\n * and the silence before anybody speaks, and ends with the silence that told us the speaker stopped. Measured\n * on a ThirdReality speaker (2026-11-03): the answer to a follow-up question — a 250 ms beep, a second of\n * digital silence, then \"Мессенджер. Да, да.\" at RMS 150–500 — came back empty from Nemotron, and Whisper\n * made \"Продолжение следует... Мессенджет. и я\" of it (the subtitle line Whisper invents for silence). Cut to\n * the speech and amplified, Whisper heard it right.\n *\n * Only what goes to the recognition changes; the diagnosis recordings keep the audio as it arrived.\n */\nimport { rms, type Bytes } from './vad';\n\n/** Audio kept before the frame where speech was detected: the soft start of a word lies below the gate. */\nexport const PRE_ROLL_MS = 400;\n/** Audio kept after the last sound: the fading end of a word. */\nexport const TAIL_MS = 400;\n/**\n * A frame in the tail counts as sound when it reaches this share of the silence threshold — low, so that a\n * quiet last syllable stays in. In a noisy room nothing is quiet enough, and the tail is simply kept.\n */\nexport const TAIL_SHARE = 0.4;\n/** Peak sample quiet speech is raised to — about half of full scale, headroom for the loudest syllable. */\nexport const TARGET_PEAK = 16000;\n/** Never amplify more than this: beyond it only the noise grows. */\nexport const MAX_GAIN = 10;\n/** Gains below this are not worth touching the audio for. */\nconst MIN_GAIN = 1.25;\n/** Never cut a recording shorter than this — a short \"да\" plus its margins. */\nconst MIN_KEEP_MS = 600;\n/** Resolution of the tail search. */\nconst FRAME_MS = 30;\n\nexport interface SttAudio {\n    /** The audio to recognise. */\n    pcm: Buffer;\n    /** Where it starts and ends in the recording, in ms. */\n    fromMs: number;\n    toMs: number;\n    /** Gain applied to it (1 = unchanged), rounded to two decimals. */\n    gain: number;\n}\n\n/**\n * Cut `pcm` (mono 16-bit) to the speech and amplify it when it is quiet.\n *\n * @param pcm the whole recording\n * @param sampleRate its sample rate\n * @param speech what the detector saw\n * @param speech.startMs where speech started, in ms (absent = keep the beginning)\n * @param speech.endLevel the silence threshold; the tail keeps everything above 40 % of it\n */\nexport function prepareForStt(\n    pcm: Bytes,\n    sampleRate: number,\n    speech: { startMs?: number; endLevel: number },\n): SttAudio {\n    const samples = Math.floor(pcm.length / 2);\n    const perMs = sampleRate / 1000;\n    const totalMs = Math.floor(samples / perMs);\n\n    let fromMs = speech.startMs !== undefined && speech.startMs >= 0 ? Math.max(0, speech.startMs - PRE_ROLL_MS) : 0;\n    fromMs = Math.min(fromMs, Math.max(0, totalMs - MIN_KEEP_MS));\n\n    // The tail: the last frame with any sound in it, plus a margin.\n    let toMs = totalMs;\n    const quiet = speech.endLevel * TAIL_SHARE;\n    if (quiet > 0) {\n        const frame = Math.max(1, Math.round(FRAME_MS * perMs));\n        const first = Math.round(fromMs * perMs);\n        let lastSound = -1;\n        for (let start = samples - frame; start >= first; start -= frame) {\n            if (rms(pcm.subarray(start * 2, (start + frame) * 2)) >= quiet) {\n                lastSound = start + frame;\n                break;\n            }\n        }\n        if (lastSound >= 0) {\n            toMs = Math.min(totalMs, Math.ceil(lastSound / perMs) + TAIL_MS);\n        }\n    }\n    if (toMs - fromMs < MIN_KEEP_MS) {\n        toMs = Math.min(totalMs, fromMs + MIN_KEEP_MS);\n    }\n\n    const from = Math.round(fromMs * perMs);\n    const to = Math.min(samples, Math.round(toMs * perMs));\n    let peak = 0;\n    for (let i = from; i < to; i++) {\n        const value = Math.abs(pcm.readInt16LE(i * 2));\n        if (value > peak) {\n            peak = value;\n        }\n    }\n    let gain = peak > 0 ? Math.min(MAX_GAIN, TARGET_PEAK / peak) : 1;\n    if (gain < MIN_GAIN) {\n        gain = 1;\n    }\n\n    const out = Buffer.alloc((to - from) * 2);\n    if (gain === 1) {\n        pcm.copy(out, 0, from * 2, to * 2);\n    } else {\n        for (let i = from; i < to; i++) {\n            const value = Math.round(pcm.readInt16LE(i * 2) * gain);\n            out.writeInt16LE(Math.max(-32768, Math.min(32767, value)), (i - from) * 2);\n        }\n    }\n    return { pcm: out, fromMs, toMs, gain: Math.round(gain * 100) / 100 };\n}\n"]}

package/build/lib/voice/vad.js

@@ -20,11 +20,16 @@
  *
  * Hence two opt-in knobs, both enabled by the ESPHome transport:
  *
- * - `skipMs` drops the leading chirp from the *analysis* (never from the audio, which still goes to STT
- *   in full — a fast speaker may already be talking).
+ * - `skipMs` drops the leading chirp from the *analysis* only. What reaches STT is cut later, from where
+ *   speech was detected minus a margin (`speechAudio.ts`) — a fast speaker may already be talking.
  * - `adaptive` derives the thresholds from a tracked noise floor instead of fixed levels. The floor
  *   follows the quietest frame seen (instant down, very slow up), so it settles on room noise even when
  *   someone starts talking immediately — which a mean or median over a calibration window would not.
+ *
+ * And one for longer requests: `longSilenceMs` after `longAfterMs` of speech. In a quiet room the silence
+ * threshold sits at its minimum, and a voice a few metres away dips below it between syllables — on a
+ * ThirdReality speaker a whole "Мессенджер" (0.7 s) stayed under 150. A pause for thought on top of that
+ * ended a long request mid-sentence, while short commands should keep closing quickly.
  */
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.SpeechDetector = void 0;
@@ -54,6 +59,8 @@ class SpeechDetector {
     fixedStart;
     fixedEnd;
     silenceMs;
+    longSilenceMs;
+    longAfterMs;
     maxMs;
     startTimeoutMs;
     skipMs;
@@ -70,6 +77,8 @@ class SpeechDetector {
     floor = null;
     /** Audio seen so far *including* the skipped lead-in, in milliseconds. */
     elapsedMs = 0;
+    /** Start of the frame that crossed {@link startLevel}, in ms from the start of the recording; -1 = not yet. */
+    speechStartMs = -1;
     /** Audio analysed so far, in milliseconds (exposed for logging). */
     totalMs = 0;
     /** Loudest frame seen so far (exposed for logging — the usual clue when a mic is mis-configured). */
@@ -79,12 +88,14 @@ class SpeechDetector {
      * it tells whether a dropped utterance was silence or a voice that was simply too quiet for the gate.
      */
     analysedPeak = 0;
-    constructor({ sampleRate = 16000, frameMs = 30, startLevel = 700, endLevel = 400, silenceMs = 900, maxMs = 12000, startTimeoutMs = 0, skipMs = 0, adaptive = false, minStartLevel = 250, minEndLevel = 150, startFactor = 3.5, endFactor = 2, warmupMs = 150, } = {}) {
+    constructor({ sampleRate = 16000, frameMs = 30, startLevel = 700, endLevel = 400, silenceMs = 900, longSilenceMs = 0, longAfterMs = 0, maxMs = 12000, startTimeoutMs = 0, skipMs = 0, adaptive = false, minStartLevel = 250, minEndLevel = 150, startFactor = 3.5, endFactor = 2, warmupMs = 150, } = {}) {
         this.frameBytes = Math.round((sampleRate * frameMs) / 1000) * 2;
         this.frameMs = frameMs;
         this.fixedStart = startLevel;
         this.fixedEnd = endLevel;
         this.silenceMs = silenceMs;
+        this.longSilenceMs = longSilenceMs;
+        this.longAfterMs = longAfterMs;
         this.maxMs = maxMs;
         this.startTimeoutMs = startTimeoutMs;
         this.skipMs = skipMs;
@@ -114,6 +125,29 @@ class SpeechDetector {
         }
         return Math.min(this.startLevel, Math.max(this.minEndLevel, (this.floor ?? 0) * this.endFactor));
     }
+    /**
+     * Silence that ends the utterance right now: {@link silenceMs}, or {@link longSilenceMs} once the speaker
+     * has talked for {@link longAfterMs} before the current pause.
+     */
+    get requiredSilenceMs() {
+        if (!this.speechSeen || !this.longAfterMs || this.longSilenceMs <= this.silenceMs) {
+            return this.silenceMs;
+        }
+        const spokenMs = this.elapsedMs - this.silentMs - this.speechStartMs;
+        return spokenMs >= this.longAfterMs ? this.longSilenceMs : this.silenceMs;
+    }
+    /** The detector's view of the recording so far. */
+    report() {
+        return {
+            peak: Math.round(this.peak),
+            analysedPeak: Math.round(this.analysedPeak),
+            noiseFloor: Math.round(this.noiseFloor),
+            startLevel: Math.round(this.startLevel),
+            endLevel: Math.round(this.endLevel),
+            speechStartMs: this.speechSeen ? this.speechStartMs : undefined,
+            silenceMs: this.requiredSilenceMs,
+        };
+    }
     /** Feed one audio chunk (any size). Returns why the utterance ended, or null to keep listening. */
     push(chunk) {
         this.pending = this.pending.length ? Buffer.concat([this.pending, chunk]) : chunk;
@@ -160,12 +194,13 @@ class SpeechDetector {
             if (!this.speechSeen) {
                 if (level >= this.startLevel) {
                     this.speechSeen = true;
+                    this.speechStartMs = this.elapsedMs - this.frameMs;
                     this.silentMs = 0;
                 }
             }
             else if (level < this.endLevel) {
                 this.silentMs += this.frameMs;
-                if (this.silentMs >= this.silenceMs) {
+                if (this.silentMs >= this.requiredSilenceMs) {
                     verdict ||= 'speech-end';
                 }
             }

package/build/lib/voice/vad.js.map

@@ -1 +1 @@
-{"version":3,"file":"vad.js","sourceRoot":"","sources":["../../../src/lib/voice/vad.ts"],"names":[],"mappings":";AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;GA0BG;;;AASH,kBAWC;AAZD,kFAAkF;AAClF,SAAgB,GAAG,CAAC,GAAU;IAC1B,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;IAC3C,IAAI,CAAC,OAAO,EAAE,CAAC;QACX,OAAO,CAAC,CAAC;IACb,CAAC;IACD,IAAI,GAAG,GAAG,CAAC,CAAC;IACZ,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,EAAE,CAAC,EAAE,EAAE,CAAC;QAC/B,MAAM,KAAK,GAAG,GAAG,CAAC,WAAW,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QACrC,GAAG,IAAI,KAAK,GAAG,KAAK,CAAC;IACzB,CAAC;IACD,OAAO,IAAI,CAAC,IAAI,CAAC,GAAG,GAAG,OAAO,CAAC,CAAC;AACpC,CAAC;AA2CD;;;;GAIG;AACH,MAAM,UAAU,GAAG,IAAI,CAAC;AAExB,MAAa,cAAc;IACN,UAAU,CAAS;IACnB,OAAO,CAAS;IAChB,UAAU,CAAS;IACnB,QAAQ,CAAS;IACjB,SAAS,CAAS;IAClB,KAAK,CAAS;IACd,cAAc,CAAS;IACvB,MAAM,CAAS;IACf,QAAQ,CAAU;IAClB,aAAa,CAAS;IACtB,WAAW,CAAS;IACpB,WAAW,CAAS;IACpB,SAAS,CAAS;IAClB,QAAQ,CAAS;IAE1B,OAAO,GAAU,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IACjC,UAAU,GAAG,KAAK,CAAC;IACnB,QAAQ,GAAG,CAAC,CAAC;IACrB,gFAAgF;IACxE,KAAK,GAAkB,IAAI,CAAC;IACpC,0EAA0E;IAClE,SAAS,GAAG,CAAC,CAAC;IACtB,oEAAoE;IACpE,OAAO,GAAG,CAAC,CAAC;IACZ,qGAAqG;IACrG,IAAI,GAAG,CAAC,CAAC;IACT;;;OAGG;IACH,YAAY,GAAG,CAAC,CAAC;IAEjB,YAAY,EACR,UAAU,GAAG,KAAK,EAClB,OAAO,GAAG,EAAE,EACZ,UAAU,GAAG,GAAG,EAChB,QAAQ,GAAG,GAAG,EACd,SAAS,GAAG,GAAG,EACf,KAAK,GAAG,KAAK,EACb,cAAc,GAAG,CAAC,EAClB,MAAM,GAAG,CAAC,EACV,QAAQ,GAAG,KAAK,EAChB,aAAa,GAAG,GAAG,EACnB,WAAW,GAAG,GAAG,EACjB,WAAW,GAAG,GAAG,EACjB,SAAS,GAAG,CAAC,EACb,QAAQ,GAAG,GAAG,MACS,EAAE;QACzB,IAAI,CAAC,UAAU,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,UAAU,GAAG,OAAO,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC;QAChE,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC;QACvB,IAAI,CAAC,UAAU,GAAG,UAAU,CAAC;QAC7B,IAAI,CAAC,QAAQ,GAAG,QAAQ,CAAC;QACzB,IAAI,CAAC,SAAS,GAAG,SAAS,CAAC;QAC3B,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,cAAc,GAAG,cAAc,CAAC;QACrC,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;QACrB,IAAI,CAAC,QAAQ,GAAG,QAAQ,CAAC;QACzB,IAAI,CAAC,aAAa,GAAG,aAAa,CAAC;QACnC,IAAI,CAAC,WAAW,GAAG,WAAW,CAAC;QAC/B,IAAI,CAAC,WAAW,GAAG,WAAW,CAAC;QAC/B,IAAI,CAAC,SAAS,GAAG,SAAS,CAAC;QAC3B,IAAI,CAAC,QAAQ,GAAG,QAAQ,CAAC;IAC7B,CAAC;IAED,kGAAkG;IAClG,IAAI,SAAS;QACT,OAAO,IAAI,CAAC,UAAU,CAAC;IAC3B,CAAC;IAED,0FAA0F;IAC1F,IAAI,UAAU;QACV,OAAO,IAAI,CAAC,KAAK,IAAI,CAAC,CAAC;IAC3B,CAAC;IAED,0FAA0F;IAC1F,IAAI,UAAU;QACV,OAAO,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,aAAa,EAAE,CAAC,IAAI,CAAC,KAAK,IAAI,CAAC,CAAC,GAAG,IAAI,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,UAAU,CAAC;IAChH,CAAC;IAED,mGAAmG;IACnG,IAAI,QAAQ;QACR,IAAI,CAAC,IAAI,CAAC,QAAQ,EAAE,CAAC;YACjB,OAAO,IAAI,CAAC,QAAQ,CAAC;QACzB,CAAC;QACD,OAAO,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,UAAU,EAAE,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,IAAI,CAAC,KAAK,IAAI,CAAC,CAAC,GAAG,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC;IACrG,CAAC;IAED,mGAAmG;IACnG,IAAI,CAAC,KAAY;QACb,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,IAAI,CAAC,OAAO,EAAE,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC;QAClF,IAAI,OAAO,GAAqB,IAAI,CAAC;QAErC,OAAO,IAAI,CAAC,OAAO,CAAC,MAAM,IAAI,IAAI,CAAC,UAAU,EAAE,CAAC;YAC5C,MAAM,KAAK,GAAG,IAAI,CAAC,OAAO,CAAC,QAAQ,CAAC,CAAC,EAAE,IAAI,CAAC,UAAU,CAAC,CAAC;YACxD,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC;YACtD,IAAI,CAAC,SAAS,IAAI,IAAI,CAAC,OAAO,CAAC;YAE/B,MAAM,KAAK,GAAG,GAAG,CAAC,KAAK,CAAC,CAAC;YACzB,IAAI,KAAK,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;gBACpB,IAAI,CAAC,IAAI,GAAG,KAAK,CAAC;YACtB,CAAC;YAED,oFAAoF;YACpF,sFAAsF;YACtF,IAAI,IAAI,CAAC,SAAS,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC;gBAChC,SAAS;YACb,CAAC;YACD,IAAI,CAAC,OAAO,IAAI,IAAI,CAAC,OAAO,CAAC;YAC7B,IAAI,KAAK,GAAG,IAAI,CAAC,YAAY,EAAE,CAAC;gBAC5B,IAAI,CAAC,YAAY,GAAG,KAAK,CAAC;YAC9B,CAAC;YAED,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAC;gBAChB,0EAA0E;gBAC1E,mFAAmF;gBACnF,oFAAoF;gBACpF,sFAAsF;gBACtF,qDAAqD;gBACrD,oEAAoE;gBACpE,mFAAmF;gBACnF,qFAAqF;gBACrF,yFAAyF;gBACzF,IAAI,CAAC,KAAK;oBACN,IAAI,CAAC,KAAK,KAAK,IAAI;wBACf,CAAC,CAAC,CAAC;wBACH,CAAC,CAAC,KAAK,GAAG,IAAI,CAAC,KAAK;4BAClB,CAAC,CAAC,KAAK;4BACP,CAAC,CAAC,IAAI,CAAC,UAAU;gCACf,CAAC,CAAC,IAAI,CAAC,KAAK;gCACZ,CAAC,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,UAAU,CAAC;gBAC7D,IAAI,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC,QAAQ,EAAE,CAAC;oBAC/B,SAAS,CAAC,2DAA2D;gBACzE,CAAC;YACL,CAAC;YAED,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,CAAC;gBACnB,IAAI,KAAK,IAAI,IAAI,CAAC,UAAU,EAAE,CAAC;oBAC3B,IAAI,CAAC,UAAU,GAAG,IAAI,CAAC;oBACvB,IAAI,CAAC,QAAQ,GAAG,CAAC,CAAC;gBACtB,CAAC;YACL,CAAC;iBAAM,IAAI,KAAK,GAAG,IAAI,CAAC,QAAQ,EAAE,CAAC;gBAC/B,IAAI,CAAC,QAAQ,IAAI,IAAI,CAAC,OAAO,CAAC;gBAC9B,IAAI,IAAI,CAAC,QAAQ,IAAI,IAAI,CAAC,SAAS,EAAE,CAAC;oBAClC,OAAO,KAAK,YAAY,CAAC;gBAC7B,CAAC;YACL,CAAC;iBAAM,CAAC;gBACJ,IAAI,CAAC,QAAQ,GAAG,CAAC,CAAC;YACtB,CAAC;YAED,IAAI,CAAC,IAAI,CAAC,UAAU,IAAI,IAAI,CAAC,cAAc,GAAG,CAAC,IAAI,IAAI,CAAC,SAAS,IAAI,IAAI,CAAC,cAAc,EAAE,CAAC;gBACvF,OAAO,KAAK,WAAW,CAAC;YAC5B,CAAC;YACD,IAAI,IAAI,CAAC,OAAO,IAAI,IAAI,CAAC,KAAK,EAAE,CAAC;gBAC7B,OAAO,KAAK,SAAS,CAAC;YAC1B,CAAC;QACL,CAAC;QACD,OAAO,OAAO,CAAC;IACnB,CAAC;CACJ;AA/JD,wCA+JC","sourcesContent":["/**\n * End-of-speech detection for satellites that stream microphone audio until they are told to stop.\n *\n * The Hannah/UDP protocol has the satellite decide when an utterance ended (`audio_end`), and Wyoming\n * clients send `audio-stop` themselves. ESPHome voice satellites do neither: they keep streaming until\n * the server sends `STT_VAD_END`/`STT_END`, so the decision has to happen here.\n *\n * This is a plain energy gate, not a neural VAD — good enough to close an utterance, and it costs\n * nothing per frame. Speech must first rise above `startLevel`; after that, `silenceMs` worth of frames\n * below `endLevel` ends it. `maxMs` is the hard stop for rooms that are simply loud.\n *\n * Fixed thresholds only work when the microphone level is known, and on a real satellite it is not.\n * Measured on a ThirdReality Voice & Music Assistant: the device plays a wake-acknowledgement chirp\n * into its own microphone that peaks at full scale (RMS ~16000, clipped), while the speech that follows\n * arrives at RMS 358–469 — below the 700 default for `startLevel` and right on the 400 for `endLevel`.\n * That breaks the gate twice over: `sawSpeech` is satisfied by the chirp rather than by any voice, so\n * the \"nothing but speech-free audio\" guard never fires, and half the real speech frames count as\n * silence, so the utterance can be cut off mid-sentence.\n *\n * Hence two opt-in knobs, both enabled by the ESPHome transport:\n *\n * - `skipMs` drops the leading chirp from the *analysis* (never from the audio, which still goes to STT\n *   in full — a fast speaker may already be talking).\n * - `adaptive` derives the thresholds from a tracked noise floor instead of fixed levels. The floor\n *   follows the quietest frame seen (instant down, very slow up), so it settles on room noise even when\n *   someone starts talking immediately — which a mean or median over a calibration window would not.\n */\n\n/**\n * A buffer that may be a view onto any `ArrayBufferLike` — what `Buffer.subarray()` and\n * `Buffer.concat()` hand back. Plain `Buffer` (= `Buffer<ArrayBuffer>`) assigns to it, not the reverse.\n */\nexport type Bytes = Buffer<ArrayBufferLike>;\n\n/** Root-mean-square amplitude (0…32767) of a mono 16-bit signed LE PCM buffer. */\nexport function rms(pcm: Bytes): number {\n    const samples = Math.floor(pcm.length / 2);\n    if (!samples) {\n        return 0;\n    }\n    let sum = 0;\n    for (let i = 0; i < samples; i++) {\n        const value = pcm.readInt16LE(i * 2);\n        sum += value * value;\n    }\n    return Math.sqrt(sum / samples);\n}\n\nexport interface SpeechDetectorOptions {\n    /** Sample rate of the incoming PCM (ESPHome satellites always capture at 16 kHz). */\n    sampleRate?: number;\n    /** Analysis window; 30 ms is the usual compromise between latency and stability. */\n    frameMs?: number;\n    /** RMS above which we consider speech to have started. Ignored when `adaptive` is on. */\n    startLevel?: number;\n    /** RMS below which a frame counts as silence. Ignored when `adaptive` is on. */\n    endLevel?: number;\n    /** Silence after speech that ends the utterance. */\n    silenceMs?: number;\n    /** Hard cap on one utterance, in case the room never falls quiet. */\n    maxMs?: number;\n    /**\n     * Give up when no speech has started this long after the stream began (the skipped lead-in included);\n     * 0 = only {@link maxMs}. A false wake would otherwise keep the microphone open for the whole `maxMs`.\n     */\n    startTimeoutMs?: number;\n    /**\n     * Leading audio to leave out of the analysis — for devices that chirp into their own microphone\n     * when the wake word fires. The audio itself is untouched; this only stops the chirp from being\n     * mistaken for speech and from poisoning the noise floor.\n     */\n    skipMs?: number;\n    /** Derive the thresholds from the tracked noise floor instead of using fixed levels. */\n    adaptive?: boolean;\n    /** Adaptive mode: `startLevel` is at least this, however quiet the room is. */\n    minStartLevel?: number;\n    /** Adaptive mode: `endLevel` is at least this. */\n    minEndLevel?: number;\n    /** Adaptive mode: speech threshold as a multiple of the noise floor. */\n    startFactor?: number;\n    /** Adaptive mode: silence threshold as a multiple of the noise floor. */\n    endFactor?: number;\n    /** Adaptive mode: settle the noise floor this long before speech may be declared. */\n    warmupMs?: number;\n}\n\n/** Why an utterance ended: the speaker stopped, we hit the hard cap, or nobody started speaking at all. */\nexport type SpeechEnd = 'speech-end' | 'timeout' | 'no-speech';\n\n/**\n * How fast the tracked noise floor may follow a *rising* ambient level, per analysed frame (30 ms).\n * Fast enough that a genuinely noisy room is matched within about a second, slow enough that a voice\n * cannot pull the floor up to its own level before it has been recognised as speech.\n */\nconst NOISE_RISE = 0.02;\n\nexport class SpeechDetector {\n    private readonly frameBytes: number;\n    private readonly frameMs: number;\n    private readonly fixedStart: number;\n    private readonly fixedEnd: number;\n    private readonly silenceMs: number;\n    private readonly maxMs: number;\n    private readonly startTimeoutMs: number;\n    private readonly skipMs: number;\n    private readonly adaptive: boolean;\n    private readonly minStartLevel: number;\n    private readonly minEndLevel: number;\n    private readonly startFactor: number;\n    private readonly endFactor: number;\n    private readonly warmupMs: number;\n\n    private pending: Bytes = Buffer.alloc(0);\n    private speechSeen = false;\n    private silentMs = 0;\n    /** Tracked room level in adaptive mode; null until the first analysed frame. */\n    private floor: number | null = null;\n    /** Audio seen so far *including* the skipped lead-in, in milliseconds. */\n    private elapsedMs = 0;\n    /** Audio analysed so far, in milliseconds (exposed for logging). */\n    totalMs = 0;\n    /** Loudest frame seen so far (exposed for logging — the usual clue when a mic is mis-configured). */\n    peak = 0;\n    /**\n     * Loudest frame after the skipped lead-in, i.e. without the wake chirp. Compared with {@link startLevel}\n     * it tells whether a dropped utterance was silence or a voice that was simply too quiet for the gate.\n     */\n    analysedPeak = 0;\n\n    constructor({\n        sampleRate = 16000,\n        frameMs = 30,\n        startLevel = 700,\n        endLevel = 400,\n        silenceMs = 900,\n        maxMs = 12000,\n        startTimeoutMs = 0,\n        skipMs = 0,\n        adaptive = false,\n        minStartLevel = 250,\n        minEndLevel = 150,\n        startFactor = 3.5,\n        endFactor = 2,\n        warmupMs = 150,\n    }: SpeechDetectorOptions = {}) {\n        this.frameBytes = Math.round((sampleRate * frameMs) / 1000) * 2;\n        this.frameMs = frameMs;\n        this.fixedStart = startLevel;\n        this.fixedEnd = endLevel;\n        this.silenceMs = silenceMs;\n        this.maxMs = maxMs;\n        this.startTimeoutMs = startTimeoutMs;\n        this.skipMs = skipMs;\n        this.adaptive = adaptive;\n        this.minStartLevel = minStartLevel;\n        this.minEndLevel = minEndLevel;\n        this.startFactor = startFactor;\n        this.endFactor = endFactor;\n        this.warmupMs = warmupMs;\n    }\n\n    /** True once any frame crossed {@link startLevel} — an utterance that never did is room noise. */\n    get sawSpeech(): boolean {\n        return this.speechSeen;\n    }\n\n    /** Tracked room level (adaptive mode), for logging. 0 before the first analysed frame. */\n    get noiseFloor(): number {\n        return this.floor ?? 0;\n    }\n\n    /** Current speech threshold — fixed, or derived from the noise floor in adaptive mode. */\n    get startLevel(): number {\n        return this.adaptive ? Math.max(this.minStartLevel, (this.floor ?? 0) * this.startFactor) : this.fixedStart;\n    }\n\n    /** Current silence threshold. Never above {@link startLevel}, so the gate keeps its hysteresis. */\n    get endLevel(): number {\n        if (!this.adaptive) {\n            return this.fixedEnd;\n        }\n        return Math.min(this.startLevel, Math.max(this.minEndLevel, (this.floor ?? 0) * this.endFactor));\n    }\n\n    /** Feed one audio chunk (any size). Returns why the utterance ended, or null to keep listening. */\n    push(chunk: Bytes): SpeechEnd | null {\n        this.pending = this.pending.length ? Buffer.concat([this.pending, chunk]) : chunk;\n        let verdict: SpeechEnd | null = null;\n\n        while (this.pending.length >= this.frameBytes) {\n            const frame = this.pending.subarray(0, this.frameBytes);\n            this.pending = this.pending.subarray(this.frameBytes);\n            this.elapsedMs += this.frameMs;\n\n            const level = rms(frame);\n            if (level > this.peak) {\n                this.peak = level;\n            }\n\n            // The device's own wake chirp: loud, clipped, and not speech. Leaving it out of the\n            // analysis is what makes `sawSpeech` mean anything, and keeps the noise floor honest.\n            if (this.elapsedMs <= this.skipMs) {\n                continue;\n            }\n            this.totalMs += this.frameMs;\n            if (level > this.analysedPeak) {\n                this.analysedPeak = level;\n            }\n\n            if (this.adaptive) {\n                // Three rules, each there to defeat a specific way of getting this wrong:\n                //  - start at zero, i.e. maximally sensitive. Seeding from the first frame instead\n                //    lets someone who talks straight after the wake word install their own voice as\n                //    the room's noise floor, which shuts the gate for good — the same trap an average\n                //    or median over a calibration window falls into.\n                //  - fall instantly, so a room that went quiet is believed at once.\n                //  - rise slowly, and not at all once speech has been recognised: otherwise a long\n                //    utterance drags the floor, and with it `endLevel`, up to its own level and cuts\n                //    itself off mid-sentence. Falling stays on, so the pauses between words still count.\n                this.floor =\n                    this.floor === null\n                        ? 0\n                        : level < this.floor\n                          ? level\n                          : this.speechSeen\n                            ? this.floor\n                            : this.floor + (level - this.floor) * NOISE_RISE;\n                if (this.totalMs < this.warmupMs) {\n                    continue; // let the floor settle before anything may count as speech\n                }\n            }\n\n            if (!this.speechSeen) {\n                if (level >= this.startLevel) {\n                    this.speechSeen = true;\n                    this.silentMs = 0;\n                }\n            } else if (level < this.endLevel) {\n                this.silentMs += this.frameMs;\n                if (this.silentMs >= this.silenceMs) {\n                    verdict ||= 'speech-end';\n                }\n            } else {\n                this.silentMs = 0;\n            }\n\n            if (!this.speechSeen && this.startTimeoutMs > 0 && this.elapsedMs >= this.startTimeoutMs) {\n                verdict ||= 'no-speech';\n            }\n            if (this.totalMs >= this.maxMs) {\n                verdict ||= 'timeout';\n            }\n        }\n        return verdict;\n    }\n}\n"]}
+{"version":3,"file":"vad.js","sourceRoot":"","sources":["../../../src/lib/voice/vad.ts"],"names":[],"mappings":";AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+BG;;;AASH,kBAWC;AAZD,kFAAkF;AAClF,SAAgB,GAAG,CAAC,GAAU;IAC1B,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;IAC3C,IAAI,CAAC,OAAO,EAAE,CAAC;QACX,OAAO,CAAC,CAAC;IACb,CAAC;IACD,IAAI,GAAG,GAAG,CAAC,CAAC;IACZ,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,EAAE,CAAC,EAAE,EAAE,CAAC;QAC/B,MAAM,KAAK,GAAG,GAAG,CAAC,WAAW,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QACrC,GAAG,IAAI,KAAK,GAAG,KAAK,CAAC;IACzB,CAAC;IACD,OAAO,IAAI,CAAC,IAAI,CAAC,GAAG,GAAG,OAAO,CAAC,CAAC;AACpC,CAAC;AAwED;;;;GAIG;AACH,MAAM,UAAU,GAAG,IAAI,CAAC;AAExB,MAAa,cAAc;IACN,UAAU,CAAS;IACnB,OAAO,CAAS;IAChB,UAAU,CAAS;IACnB,QAAQ,CAAS;IACjB,SAAS,CAAS;IAClB,aAAa,CAAS;IACtB,WAAW,CAAS;IACpB,KAAK,CAAS;IACd,cAAc,CAAS;IACvB,MAAM,CAAS;IACf,QAAQ,CAAU;IAClB,aAAa,CAAS;IACtB,WAAW,CAAS;IACpB,WAAW,CAAS;IACpB,SAAS,CAAS;IAClB,QAAQ,CAAS;IAE1B,OAAO,GAAU,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IACjC,UAAU,GAAG,KAAK,CAAC;IACnB,QAAQ,GAAG,CAAC,CAAC;IACrB,gFAAgF;IACxE,KAAK,GAAkB,IAAI,CAAC;IACpC,0EAA0E;IAClE,SAAS,GAAG,CAAC,CAAC;IACtB,+GAA+G;IAC/G,aAAa,GAAG,CAAC,CAAC,CAAC;IACnB,oEAAoE;IACpE,OAAO,GAAG,CAAC,CAAC;IACZ,qGAAqG;IACrG,IAAI,GAAG,CAAC,CAAC;IACT;;;OAGG;IACH,YAAY,GAAG,CAAC,CAAC;IAEjB,YAAY,EACR,UAAU,GAAG,KAAK,EAClB,OAAO,GAAG,EAAE,EACZ,UAAU,GAAG,GAAG,EAChB,QAAQ,GAAG,GAAG,EACd,SAAS,GAAG,GAAG,EACf,aAAa,GAAG,CAAC,EACjB,WAAW,GAAG,CAAC,EACf,KAAK,GAAG,KAAK,EACb,cAAc,GAAG,CAAC,EAClB,MAAM,GAAG,CAAC,EACV,QAAQ,GAAG,KAAK,EAChB,aAAa,GAAG,GAAG,EACnB,WAAW,GAAG,GAAG,EACjB,WAAW,GAAG,GAAG,EACjB,SAAS,GAAG,CAAC,EACb,QAAQ,GAAG,GAAG,MACS,EAAE;QACzB,IAAI,CAAC,UAAU,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,UAAU,GAAG,OAAO,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC;QAChE,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC;QACvB,IAAI,CAAC,UAAU,GAAG,UAAU,CAAC;QAC7B,IAAI,CAAC,QAAQ,GAAG,QAAQ,CAAC;QACzB,IAAI,CAAC,SAAS,GAAG,SAAS,CAAC;QAC3B,IAAI,CAAC,aAAa,GAAG,aAAa,CAAC;QACnC,IAAI,CAAC,WAAW,GAAG,WAAW,CAAC;QAC/B,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,cAAc,GAAG,cAAc,CAAC;QACrC,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;QACrB,IAAI,CAAC,QAAQ,GAAG,QAAQ,CAAC;QACzB,IAAI,CAAC,aAAa,GAAG,aAAa,CAAC;QACnC,IAAI,CAAC,WAAW,GAAG,WAAW,CAAC;QAC/B,IAAI,CAAC,WAAW,GAAG,WAAW,CAAC;QAC/B,IAAI,CAAC,SAAS,GAAG,SAAS,CAAC;QAC3B,IAAI,CAAC,QAAQ,GAAG,QAAQ,CAAC;IAC7B,CAAC;IAED,kGAAkG;IAClG,IAAI,SAAS;QACT,OAAO,IAAI,CAAC,UAAU,CAAC;IAC3B,CAAC;IAED,0FAA0F;IAC1F,IAAI,UAAU;QACV,OAAO,IAAI,CAAC,KAAK,IAAI,CAAC,CAAC;IAC3B,CAAC;IAED,0FAA0F;IAC1F,IAAI,UAAU;QACV,OAAO,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,aAAa,EAAE,CAAC,IAAI,CAAC,KAAK,IAAI,CAAC,CAAC,GAAG,IAAI,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,UAAU,CAAC;IAChH,CAAC;IAED,mGAAmG;IACnG,IAAI,QAAQ;QACR,IAAI,CAAC,IAAI,CAAC,QAAQ,EAAE,CAAC;YACjB,OAAO,IAAI,CAAC,QAAQ,CAAC;QACzB,CAAC;QACD,OAAO,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,UAAU,EAAE,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,IAAI,CAAC,KAAK,IAAI,CAAC,CAAC,GAAG,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC;IACrG,CAAC;IAED;;;OAGG;IACH,IAAI,iBAAiB;QACjB,IAAI,CAAC,IAAI,CAAC,UAAU,IAAI,CAAC,IAAI,CAAC,WAAW,IAAI,IAAI,CAAC,aAAa,IAAI,IAAI,CAAC,SAAS,EAAE,CAAC;YAChF,OAAO,IAAI,CAAC,SAAS,CAAC;QAC1B,CAAC;QACD,MAAM,QAAQ,GAAG,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,QAAQ,GAAG,IAAI,CAAC,aAAa,CAAC;QACrE,OAAO,QAAQ,IAAI,IAAI,CAAC,WAAW,CAAC,CAAC,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC;IAC9E,CAAC;IAED,mDAAmD;IACnD,MAAM;QACF,OAAO;YACH,IAAI,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC;YAC3B,YAAY,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,YAAY,CAAC;YAC3C,UAAU,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,UAAU,CAAC;YACvC,UAAU,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,UAAU,CAAC;YACvC,QAAQ,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,QAAQ,CAAC;YACnC,aAAa,EAAE,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,SAAS;YAC/D,SAAS,EAAE,IAAI,CAAC,iBAAiB;SACpC,CAAC;IACN,CAAC;IAED,mGAAmG;IACnG,IAAI,CAAC,KAAY;QACb,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,IAAI,CAAC,OAAO,EAAE,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC;QAClF,IAAI,OAAO,GAAqB,IAAI,CAAC;QAErC,OAAO,IAAI,CAAC,OAAO,CAAC,MAAM,IAAI,IAAI,CAAC,UAAU,EAAE,CAAC;YAC5C,MAAM,KAAK,GAAG,IAAI,CAAC,OAAO,CAAC,QAAQ,CAAC,CAAC,EAAE,IAAI,CAAC,UAAU,CAAC,CAAC;YACxD,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC;YACtD,IAAI,CAAC,SAAS,IAAI,IAAI,CAAC,OAAO,CAAC;YAE/B,MAAM,KAAK,GAAG,GAAG,CAAC,KAAK,CAAC,CAAC;YACzB,IAAI,KAAK,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;gBACpB,IAAI,CAAC,IAAI,GAAG,KAAK,CAAC;YACtB,CAAC;YAED,oFAAoF;YACpF,sFAAsF;YACtF,IAAI,IAAI,CAAC,SAAS,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC;gBAChC,SAAS;YACb,CAAC;YACD,IAAI,CAAC,OAAO,IAAI,IAAI,CAAC,OAAO,CAAC;YAC7B,IAAI,KAAK,GAAG,IAAI,CAAC,YAAY,EAAE,CAAC;gBAC5B,IAAI,CAAC,YAAY,GAAG,KAAK,CAAC;YAC9B,CAAC;YAED,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAC;gBAChB,0EAA0E;gBAC1E,mFAAmF;gBACnF,oFAAoF;gBACpF,sFAAsF;gBACtF,qDAAqD;gBACrD,oEAAoE;gBACpE,mFAAmF;gBACnF,qFAAqF;gBACrF,yFAAyF;gBACzF,IAAI,CAAC,KAAK;oBACN,IAAI,CAAC,KAAK,KAAK,IAAI;wBACf,CAAC,CAAC,CAAC;wBACH,CAAC,CAAC,KAAK,GAAG,IAAI,CAAC,KAAK;4BAClB,CAAC,CAAC,KAAK;4BACP,CAAC,CAAC,IAAI,CAAC,UAAU;gCACf,CAAC,CAAC,IAAI,CAAC,KAAK;gCACZ,CAAC,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,UAAU,CAAC;gBAC7D,IAAI,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC,QAAQ,EAAE,CAAC;oBAC/B,SAAS,CAAC,2DAA2D;gBACzE,CAAC;YACL,CAAC;YAED,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,CAAC;gBACnB,IAAI,KAAK,IAAI,IAAI,CAAC,UAAU,EAAE,CAAC;oBAC3B,IAAI,CAAC,UAAU,GAAG,IAAI,CAAC;oBACvB,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,OAAO,CAAC;oBACnD,IAAI,CAAC,QAAQ,GAAG,CAAC,CAAC;gBACtB,CAAC;YACL,CAAC;iBAAM,IAAI,KAAK,GAAG,IAAI,CAAC,QAAQ,EAAE,CAAC;gBAC/B,IAAI,CAAC,QAAQ,IAAI,IAAI,CAAC,OAAO,CAAC;gBAC9B,IAAI,IAAI,CAAC,QAAQ,IAAI,IAAI,CAAC,iBAAiB,EAAE,CAAC;oBAC1C,OAAO,KAAK,YAAY,CAAC;gBAC7B,CAAC;YACL,CAAC;iBAAM,CAAC;gBACJ,IAAI,CAAC,QAAQ,GAAG,CAAC,CAAC;YACtB,CAAC;YAED,IAAI,CAAC,IAAI,CAAC,UAAU,IAAI,IAAI,CAAC,cAAc,GAAG,CAAC,IAAI,IAAI,CAAC,SAAS,IAAI,IAAI,CAAC,cAAc,EAAE,CAAC;gBACvF,OAAO,KAAK,WAAW,CAAC;YAC5B,CAAC;YACD,IAAI,IAAI,CAAC,OAAO,IAAI,IAAI,CAAC,KAAK,EAAE,CAAC;gBAC7B,OAAO,KAAK,SAAS,CAAC;YAC1B,CAAC;QACL,CAAC;QACD,OAAO,OAAO,CAAC;IACnB,CAAC;CACJ;AAjMD,wCAiMC","sourcesContent":["/**\n * End-of-speech detection for satellites that stream microphone audio until they are told to stop.\n *\n * The Hannah/UDP protocol has the satellite decide when an utterance ended (`audio_end`), and Wyoming\n * clients send `audio-stop` themselves. ESPHome voice satellites do neither: they keep streaming until\n * the server sends `STT_VAD_END`/`STT_END`, so the decision has to happen here.\n *\n * This is a plain energy gate, not a neural VAD — good enough to close an utterance, and it costs\n * nothing per frame. Speech must first rise above `startLevel`; after that, `silenceMs` worth of frames\n * below `endLevel` ends it. `maxMs` is the hard stop for rooms that are simply loud.\n *\n * Fixed thresholds only work when the microphone level is known, and on a real satellite it is not.\n * Measured on a ThirdReality Voice & Music Assistant: the device plays a wake-acknowledgement chirp\n * into its own microphone that peaks at full scale (RMS ~16000, clipped), while the speech that follows\n * arrives at RMS 358–469 — below the 700 default for `startLevel` and right on the 400 for `endLevel`.\n * That breaks the gate twice over: `sawSpeech` is satisfied by the chirp rather than by any voice, so\n * the \"nothing but speech-free audio\" guard never fires, and half the real speech frames count as\n * silence, so the utterance can be cut off mid-sentence.\n *\n * Hence two opt-in knobs, both enabled by the ESPHome transport:\n *\n * - `skipMs` drops the leading chirp from the *analysis* only. What reaches STT is cut later, from where\n *   speech was detected minus a margin (`speechAudio.ts`) — a fast speaker may already be talking.\n * - `adaptive` derives the thresholds from a tracked noise floor instead of fixed levels. The floor\n *   follows the quietest frame seen (instant down, very slow up), so it settles on room noise even when\n *   someone starts talking immediately — which a mean or median over a calibration window would not.\n *\n * And one for longer requests: `longSilenceMs` after `longAfterMs` of speech. In a quiet room the silence\n * threshold sits at its minimum, and a voice a few metres away dips below it between syllables — on a\n * ThirdReality speaker a whole \"Мессенджер\" (0.7 s) stayed under 150. A pause for thought on top of that\n * ended a long request mid-sentence, while short commands should keep closing quickly.\n */\n\n/**\n * A buffer that may be a view onto any `ArrayBufferLike` — what `Buffer.subarray()` and\n * `Buffer.concat()` hand back. Plain `Buffer` (= `Buffer<ArrayBuffer>`) assigns to it, not the reverse.\n */\nexport type Bytes = Buffer<ArrayBufferLike>;\n\n/** Root-mean-square amplitude (0…32767) of a mono 16-bit signed LE PCM buffer. */\nexport function rms(pcm: Bytes): number {\n    const samples = Math.floor(pcm.length / 2);\n    if (!samples) {\n        return 0;\n    }\n    let sum = 0;\n    for (let i = 0; i < samples; i++) {\n        const value = pcm.readInt16LE(i * 2);\n        sum += value * value;\n    }\n    return Math.sqrt(sum / samples);\n}\n\nexport interface SpeechDetectorOptions {\n    /** Sample rate of the incoming PCM (ESPHome satellites always capture at 16 kHz). */\n    sampleRate?: number;\n    /** Analysis window; 30 ms is the usual compromise between latency and stability. */\n    frameMs?: number;\n    /** RMS above which we consider speech to have started. Ignored when `adaptive` is on. */\n    startLevel?: number;\n    /** RMS below which a frame counts as silence. Ignored when `adaptive` is on. */\n    endLevel?: number;\n    /** Silence after speech that ends the utterance. */\n    silenceMs?: number;\n    /**\n     * Silence that ends the utterance once the speaker has talked for {@link longAfterMs} (0 = always\n     * {@link silenceMs}). Somebody dictating a longer request pauses to think, and a short command should\n     * still close fast.\n     */\n    longSilenceMs?: number;\n    /** Speech (from its start to the pause) after which {@link longSilenceMs} applies; 0 = never. */\n    longAfterMs?: number;\n    /** Hard cap on one utterance, in case the room never falls quiet. */\n    maxMs?: number;\n    /**\n     * Give up when no speech has started this long after the stream began (the skipped lead-in included);\n     * 0 = only {@link maxMs}. A false wake would otherwise keep the microphone open for the whole `maxMs`.\n     */\n    startTimeoutMs?: number;\n    /**\n     * Leading audio to leave out of the analysis — for devices that chirp into their own microphone\n     * when the wake word fires. The audio itself is untouched; this only stops the chirp from being\n     * mistaken for speech and from poisoning the noise floor.\n     */\n    skipMs?: number;\n    /** Derive the thresholds from the tracked noise floor instead of using fixed levels. */\n    adaptive?: boolean;\n    /** Adaptive mode: `startLevel` is at least this, however quiet the room is. */\n    minStartLevel?: number;\n    /** Adaptive mode: `endLevel` is at least this. */\n    minEndLevel?: number;\n    /** Adaptive mode: speech threshold as a multiple of the noise floor. */\n    startFactor?: number;\n    /** Adaptive mode: silence threshold as a multiple of the noise floor. */\n    endFactor?: number;\n    /** Adaptive mode: settle the noise floor this long before speech may be declared. */\n    warmupMs?: number;\n}\n\n/** Why an utterance ended: the speaker stopped, we hit the hard cap, or nobody started speaking at all. */\nexport type SpeechEnd = 'speech-end' | 'timeout' | 'no-speech';\n\n/** What the detector made of one recording — for the request trace and the diagnosis recordings. */\nexport interface VadReport {\n    /** Loudest frame, the wake chirp included. */\n    peak: number;\n    /** Loudest frame after the skipped lead-in. */\n    analysedPeak: number;\n    noiseFloor: number;\n    startLevel: number;\n    endLevel: number;\n    /** Where speech began, in ms from the start of the recording; absent when it never did. */\n    speechStartMs?: number;\n    /** Why the recording ended; 'stream' = the device stopped streaming by itself. */\n    end?: SpeechEnd | 'stream';\n    /** Silence that was needed to end it at that point (longer once the speaker talked for a while). */\n    silenceMs?: number;\n    /** RMS per 100 ms of the whole recording. */\n    levels?: number[];\n    /** What went to speech recognition: the recording cut to the speech, and the gain applied to it. */\n    stt?: { fromMs: number; toMs: number; gain: number };\n}\n\n/**\n * How fast the tracked noise floor may follow a *rising* ambient level, per analysed frame (30 ms).\n * Fast enough that a genuinely noisy room is matched within about a second, slow enough that a voice\n * cannot pull the floor up to its own level before it has been recognised as speech.\n */\nconst NOISE_RISE = 0.02;\n\nexport class SpeechDetector {\n    private readonly frameBytes: number;\n    private readonly frameMs: number;\n    private readonly fixedStart: number;\n    private readonly fixedEnd: number;\n    private readonly silenceMs: number;\n    private readonly longSilenceMs: number;\n    private readonly longAfterMs: number;\n    private readonly maxMs: number;\n    private readonly startTimeoutMs: number;\n    private readonly skipMs: number;\n    private readonly adaptive: boolean;\n    private readonly minStartLevel: number;\n    private readonly minEndLevel: number;\n    private readonly startFactor: number;\n    private readonly endFactor: number;\n    private readonly warmupMs: number;\n\n    private pending: Bytes = Buffer.alloc(0);\n    private speechSeen = false;\n    private silentMs = 0;\n    /** Tracked room level in adaptive mode; null until the first analysed frame. */\n    private floor: number | null = null;\n    /** Audio seen so far *including* the skipped lead-in, in milliseconds. */\n    private elapsedMs = 0;\n    /** Start of the frame that crossed {@link startLevel}, in ms from the start of the recording; -1 = not yet. */\n    speechStartMs = -1;\n    /** Audio analysed so far, in milliseconds (exposed for logging). */\n    totalMs = 0;\n    /** Loudest frame seen so far (exposed for logging — the usual clue when a mic is mis-configured). */\n    peak = 0;\n    /**\n     * Loudest frame after the skipped lead-in, i.e. without the wake chirp. Compared with {@link startLevel}\n     * it tells whether a dropped utterance was silence or a voice that was simply too quiet for the gate.\n     */\n    analysedPeak = 0;\n\n    constructor({\n        sampleRate = 16000,\n        frameMs = 30,\n        startLevel = 700,\n        endLevel = 400,\n        silenceMs = 900,\n        longSilenceMs = 0,\n        longAfterMs = 0,\n        maxMs = 12000,\n        startTimeoutMs = 0,\n        skipMs = 0,\n        adaptive = false,\n        minStartLevel = 250,\n        minEndLevel = 150,\n        startFactor = 3.5,\n        endFactor = 2,\n        warmupMs = 150,\n    }: SpeechDetectorOptions = {}) {\n        this.frameBytes = Math.round((sampleRate * frameMs) / 1000) * 2;\n        this.frameMs = frameMs;\n        this.fixedStart = startLevel;\n        this.fixedEnd = endLevel;\n        this.silenceMs = silenceMs;\n        this.longSilenceMs = longSilenceMs;\n        this.longAfterMs = longAfterMs;\n        this.maxMs = maxMs;\n        this.startTimeoutMs = startTimeoutMs;\n        this.skipMs = skipMs;\n        this.adaptive = adaptive;\n        this.minStartLevel = minStartLevel;\n        this.minEndLevel = minEndLevel;\n        this.startFactor = startFactor;\n        this.endFactor = endFactor;\n        this.warmupMs = warmupMs;\n    }\n\n    /** True once any frame crossed {@link startLevel} — an utterance that never did is room noise. */\n    get sawSpeech(): boolean {\n        return this.speechSeen;\n    }\n\n    /** Tracked room level (adaptive mode), for logging. 0 before the first analysed frame. */\n    get noiseFloor(): number {\n        return this.floor ?? 0;\n    }\n\n    /** Current speech threshold — fixed, or derived from the noise floor in adaptive mode. */\n    get startLevel(): number {\n        return this.adaptive ? Math.max(this.minStartLevel, (this.floor ?? 0) * this.startFactor) : this.fixedStart;\n    }\n\n    /** Current silence threshold. Never above {@link startLevel}, so the gate keeps its hysteresis. */\n    get endLevel(): number {\n        if (!this.adaptive) {\n            return this.fixedEnd;\n        }\n        return Math.min(this.startLevel, Math.max(this.minEndLevel, (this.floor ?? 0) * this.endFactor));\n    }\n\n    /**\n     * Silence that ends the utterance right now: {@link silenceMs}, or {@link longSilenceMs} once the speaker\n     * has talked for {@link longAfterMs} before the current pause.\n     */\n    get requiredSilenceMs(): number {\n        if (!this.speechSeen || !this.longAfterMs || this.longSilenceMs <= this.silenceMs) {\n            return this.silenceMs;\n        }\n        const spokenMs = this.elapsedMs - this.silentMs - this.speechStartMs;\n        return spokenMs >= this.longAfterMs ? this.longSilenceMs : this.silenceMs;\n    }\n\n    /** The detector's view of the recording so far. */\n    report(): VadReport {\n        return {\n            peak: Math.round(this.peak),\n            analysedPeak: Math.round(this.analysedPeak),\n            noiseFloor: Math.round(this.noiseFloor),\n            startLevel: Math.round(this.startLevel),\n            endLevel: Math.round(this.endLevel),\n            speechStartMs: this.speechSeen ? this.speechStartMs : undefined,\n            silenceMs: this.requiredSilenceMs,\n        };\n    }\n\n    /** Feed one audio chunk (any size). Returns why the utterance ended, or null to keep listening. */\n    push(chunk: Bytes): SpeechEnd | null {\n        this.pending = this.pending.length ? Buffer.concat([this.pending, chunk]) : chunk;\n        let verdict: SpeechEnd | null = null;\n\n        while (this.pending.length >= this.frameBytes) {\n            const frame = this.pending.subarray(0, this.frameBytes);\n            this.pending = this.pending.subarray(this.frameBytes);\n            this.elapsedMs += this.frameMs;\n\n            const level = rms(frame);\n            if (level > this.peak) {\n                this.peak = level;\n            }\n\n            // The device's own wake chirp: loud, clipped, and not speech. Leaving it out of the\n            // analysis is what makes `sawSpeech` mean anything, and keeps the noise floor honest.\n            if (this.elapsedMs <= this.skipMs) {\n                continue;\n            }\n            this.totalMs += this.frameMs;\n            if (level > this.analysedPeak) {\n                this.analysedPeak = level;\n            }\n\n            if (this.adaptive) {\n                // Three rules, each there to defeat a specific way of getting this wrong:\n                //  - start at zero, i.e. maximally sensitive. Seeding from the first frame instead\n                //    lets someone who talks straight after the wake word install their own voice as\n                //    the room's noise floor, which shuts the gate for good — the same trap an average\n                //    or median over a calibration window falls into.\n                //  - fall instantly, so a room that went quiet is believed at once.\n                //  - rise slowly, and not at all once speech has been recognised: otherwise a long\n                //    utterance drags the floor, and with it `endLevel`, up to its own level and cuts\n                //    itself off mid-sentence. Falling stays on, so the pauses between words still count.\n                this.floor =\n                    this.floor === null\n                        ? 0\n                        : level < this.floor\n                          ? level\n                          : this.speechSeen\n                            ? this.floor\n                            : this.floor + (level - this.floor) * NOISE_RISE;\n                if (this.totalMs < this.warmupMs) {\n                    continue; // let the floor settle before anything may count as speech\n                }\n            }\n\n            if (!this.speechSeen) {\n                if (level >= this.startLevel) {\n                    this.speechSeen = true;\n                    this.speechStartMs = this.elapsedMs - this.frameMs;\n                    this.silentMs = 0;\n                }\n            } else if (level < this.endLevel) {\n                this.silentMs += this.frameMs;\n                if (this.silentMs >= this.requiredSilenceMs) {\n                    verdict ||= 'speech-end';\n                }\n            } else {\n                this.silentMs = 0;\n            }\n\n            if (!this.speechSeen && this.startTimeoutMs > 0 && this.elapsedMs >= this.startTimeoutMs) {\n                verdict ||= 'no-speech';\n            }\n            if (this.totalMs >= this.maxMs) {\n                verdict ||= 'timeout';\n            }\n        }\n        return verdict;\n    }\n}\n"]}

package/build/main.js

@@ -1361,6 +1361,11 @@ class Assistant extends adapter_core_1.Adapter {
                             unit: entity.unit || undefined,
                         };
                 await this.setObjectNotExistsAsync(id, { type: 'state', common, native: {} });
+                if (entity.kind === 'number') {
+                    // Objects from before float32() kept the raw float noise (step 0.0010000000474974513);
+                    // the device may also change its range with a firmware update.
+                    await this.syncNumberRange(id, entity);
+                }
                 this.subscribeStates(id);
             }
         }
@@ -1399,6 +1404,18 @@ class Assistant extends adapter_core_1.Adapter {
         await ready;
         await this.writeSatelliteControl(satId, entity);
     }
+    /** Bring min/max/step of an existing number control in line with what the device announces now. */
+    async syncNumberRange(id, entity) {
+        const obj = await this.getObjectAsync(id);
+        if (!obj || obj.type !== 'state') {
+            return;
+        }
+        const wanted = { min: entity.min, max: entity.max, step: entity.step };
+        const common = obj.common;
+        if (common.min !== wanted.min || common.max !== wanted.max || common.step !== wanted.step) {
+            await this.extendObjectAsync(id, { common: wanted });
+        }
+    }
     /** Write one entity's current value into its state(s). Always acked — this is the device talking. */
     async writeSatelliteControl(satId, entity) {
         const base = `satellites.${satId}.controls.${entity.objectId}`;
@@ -1416,7 +1433,15 @@ class Assistant extends adapter_core_1.Adapter {
             await this.setStateAsync(`${base}.progress`, { val: v.progress, ack: true });
         }
         else {
-            await this.setStateAsync(base, { val: entity.value, ack: true });
+            let val = entity.value;
+            // The device says how many decimals of a measurement mean something; the rest is noise.
+            if (entity.kind === 'sensor' &&
+                typeof val === 'number' &&
+                entity.decimals !== undefined &&
+                entity.decimals >= 0) {
+                val = Number(val.toFixed(Math.min(entity.decimals, 10)));
+            }
+            await this.setStateAsync(base, { val, ack: true });
         }
     }
     /**