@reins-ai/voice 0.0.0-stage → 0.1.0

package/dist/ear.d.ts

@@ -0,0 +1,158 @@
+import { type Log } from './log.js';
+import type { OpenMic } from './mic.js';
+import type { SpeechCheck } from './speech-check.js';
+import type { Transcriber } from './stt.js';
+import { type TurnCheck } from './turn.js';
+import { type VadOptions } from './vad.js';
+export declare const MAX_WAIT_MS = 60000;
+/**
+ * What the AI said may come back through the microphone (echo) up to this long after it finished saying it.
+ */
+export declare const ECHO_WINDOW_MS = 10000;
+/** A sentence is echo if at least this share of its words came from what the AI said… */
+export declare const ECHO_OVERLAP = 0.7;
+/** …and it has at least this many words (a user's "yes" or "ok" is never echo). */
+export declare const ECHO_MIN_WORDS = 3;
+/**
+ * If the AI does not call listen again within this time, the conversation is over and the microphone closes.
+ */
+export declare const IDLE_CLOSE_MS = 30000;
+/** A "sentence" with less voice energy than this is a knock or a noise and is not transcribed. */
+export declare const MIN_VOICED_MS = 300;
+/**
+ * With something already understood, listen waits at most this long for the speaker to stop (it may be a TV).
+ */
+export declare const SETTLE_MS = 3000;
+/**
+ * At every 0.3 s pause (10 blocks) Smart Turn is asked whether the user has finished. If not, the 1.5 s limit
+ * applies.
+ */
+export declare const PAUSE_FRAMES = 10;
+/**
+ * Smart Turn only closes a sentence early when it has at least this much voice, so it does not split a short
+ * question in pieces.
+ */
+export declare const MIN_TURN_VOICED_MS = 1000;
+/**
+ * A sentence that started while the AI was talking and is not its echo only cuts the AI off with at least
+ * this much voice. Echo leftovers are chunks with little voice (330 ms in 3 s) that Whisper sometimes
+ * transcribes as other words; a user saying "stop, stop" had 750 ms.
+ */
+export declare const MIN_INTERRUPT_VOICED_MS = 500;
+/**
+ * …and, under 3 words, it must be almost all voice. An echo leftover is a few words in a long chunk with
+ * little voice (510 ms of voice in 3.4 s, 15 %); a single "stop" from the user is almost all voice.
+ */
+export declare const MIN_INTERRUPT_VOICE_RATIO = 0.3;
+/**
+ * What listen returns when the user cut the AI off, so the AI knows the user did not hear everything it said.
+ */
+export declare const INTERRUPTED = "(the user interrupted you while you were speaking)";
+export interface EarOptions {
+    openMic: OpenMic;
+    transcriber: Transcriber;
+    /**
+     * Checks that a sentence has human voice before Whisper: over noise, Whisper makes up sentences (from the
+     * subtitles it was trained on). Works for any language.
+     */
+    speechCheck?: SpeechCheck;
+    vad?: VadOptions;
+    /**
+     * Decides whether a pause ends the turn (Smart Turn). With it, the sentence closes after 0.3 s of silence
+     * if the user has finished; without it, always after 1.5 s.
+     */
+    turnCheck?: TurnCheck;
+    /** A user sentence just closed while the AI listens, to let the user know they were heard. */
+    onPhrase?: () => void;
+    /** The user started talking while the AI talks: it must go quiet. */
+    onBargeIn?: () => void;
+    log?: Log;
+}
+/**
+ * Continuous listening. The microphone stays open from the first listen, also while the AI talks, so the user
+ * can answer or interrupt at once. Each long pause closes a sentence, which is transcribed and stored; listen
+ * returns it whole. Whatever matches what the AI just said is its own echo from the speakers and is
+ * discarded.
+ */
+export declare class Ear {
+    private readonly options;
+    private mic;
+    private vad;
+    private leftover;
+    /**
+     * Understood text not collected by listen yet. `fromUser`: it started while the AI was quiet, so it is not
+     * an echo leftover.
+     */
+    private heard;
+    private failure;
+    private waiters;
+    private transcribing;
+    private pending;
+    private speaking;
+    private said;
+    /** Bumped on close, so transcriptions in progress are discarded. */
+    private generation;
+    private idleTimer;
+    private heardAt;
+    /** Bumped on every interrupt, so the waiting listen returns. */
+    private interrupts;
+    /** Bumped on every closed sentence, so a Smart Turn answer for a previous sentence is ignored. */
+    private utteranceId;
+    /** Smart Turn is deciding: it is not asked again, since each pass delays Silero and Whisper. */
+    private checkingTurn;
+    private gate;
+    /** The current sentence started while the AI was talking, so it is almost always its echo. */
+    private startedWithAI;
+    /** The user cut the AI off; the AI is told on the next listen. */
+    private interruptedAI;
+    constructor(options: EarOptions);
+    /**
+     * Everything the user said since the last call, whole: waits for them to stop talking and for the rest to
+     * be transcribed. It never gets stuck with endless background voice (a TV, a video): with something already
+     * understood, it returns at most SETTLE_MS later, and after maxWaitMs it returns whatever there is (or
+     * nothing). Anything still playing is not lost: it goes to the next listen.
+     */
+    next(maxWaitMs?: number): Promise<string>;
+    /**
+     * A 30 ms chunk about to play from the speaker (its level and when it plays). `first` starts a sentence.
+     */
+    playing(level: number, at: number, first?: boolean): void;
+    /**
+     * The AI is about to say this text. Returns the function to call when it finishes: it closes the current
+     * sentence, which is its echo from the speakers, so it does not merge with what the user says next (if they
+     * merged, the echo filter would also discard the user's words).
+     */
+    saying(text: string): () => void;
+    /**
+     * Another tool arrives (speak, stop or another listen): the waiting listen returns now, without consuming
+     * what was heard. This way the AI can speak even if an earlier listen is still waiting.
+     */
+    interrupt(): void;
+    /** Ends the conversation: closes the microphone and clears everything pending. */
+    stop(): void;
+    private open;
+    private onAudio;
+    /**
+     * The user talks over the AI: it goes quiet, and the user's sentence starts here (what came before was the
+     * AI's voice, which the echo filter would discard along with the user's words).
+     */
+    private bargeIn;
+    /**
+     * 0.3 s pause: Smart Turn tells whether the user has finished. If so and they are still quiet, the sentence
+     * closes now instead of after 1.5 s. If they started talking again meanwhile, the answer is ignored.
+     */
+    private checkTurn;
+    /** Closes the current sentence and prepares the next one; the detector keeps the room noise level. */
+    private endUtterance;
+    private log;
+    private aiSpeaking;
+    /** Sentences are transcribed in order, one after another. */
+    private transcribe;
+    /**
+     * With fewer than `minWords` words it is never echo (a user's "yes"); 1 for what started while the AI was
+     * talking.
+     */
+    private isEcho;
+    private wake;
+}
+//# sourceMappingURL=ear.d.ts.map

package/dist/ear.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"ear.d.ts","sourceRoot":"","sources":["../src/ear.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,KAAK,GAAG,EAAU,MAAM,UAAU,CAAC;AAC5C,OAAO,KAAK,EAAO,OAAO,EAAE,MAAM,UAAU,CAAC;AAC7C,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AACrD,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,UAAU,CAAC;AAE5C,OAAO,EAAiB,KAAK,SAAS,EAAE,MAAM,WAAW,CAAC;AAC1D,OAAO,EAAiB,KAAK,UAAU,EAAiB,MAAM,UAAU,CAAC;AAEzE,eAAO,MAAM,WAAW,QAAS,CAAC;AAClC;;GAEG;AACH,eAAO,MAAM,cAAc,QAAS,CAAC;AACrC,yFAAyF;AACzF,eAAO,MAAM,YAAY,MAAM,CAAC;AAChC,mFAAmF;AACnF,eAAO,MAAM,cAAc,IAAI,CAAC;AAChC;;GAEG;AACH,eAAO,MAAM,aAAa,QAAS,CAAC;AACpC,kGAAkG;AAClG,eAAO,MAAM,aAAa,MAAM,CAAC;AACjC;;GAEG;AACH,eAAO,MAAM,SAAS,OAAQ,CAAC;AAC/B;;;GAGG;AACH,eAAO,MAAM,YAAY,KAAK,CAAC;AAC/B;;;GAGG;AACH,eAAO,MAAM,kBAAkB,OAAQ,CAAC;AAMxC;;;;GAIG;AACH,eAAO,MAAM,uBAAuB,MAAM,CAAC;AAC3C;;;GAGG;AACH,eAAO,MAAM,yBAAyB,MAAM,CAAC;AAS7C;;GAEG;AACH,eAAO,MAAM,WAAW,uDAAuD,CAAC;AAchF,MAAM,WAAW,UAAU;IACzB,OAAO,EAAE,OAAO,CAAC;IACjB,WAAW,EAAE,WAAW,CAAC;IACzB;;;OAGG;IACH,WAAW,CAAC,EAAE,WAAW,CAAC;IAC1B,GAAG,CAAC,EAAE,UAAU,CAAC;IACjB;;;OAGG;IACH,SAAS,CAAC,EAAE,SAAS,CAAC;IACtB,8FAA8F;IAC9F,QAAQ,CAAC,EAAE,MAAM,IAAI,CAAC;IACtB,qEAAqE;IACrE,SAAS,CAAC,EAAE,MAAM,IAAI,CAAC;IACvB,GAAG,CAAC,EAAE,GAAG,CAAC;CACX;AAgBD;;;;;GAKG;AACH,qBAAa,GAAG;IA+BF,OAAO,CAAC,QAAQ,CAAC,OAAO;IA9BpC,OAAO,CAAC,GAAG,CAAkB;IAC7B,OAAO,CAAC,GAAG,CAAgB;IAC3B,OAAO,CAAC,QAAQ,CAAmB;IACnC;;;OAGG;IACH,OAAO,CAAC,KAAK,CAA6C;IAC1D,OAAO,CAAC,OAAO,CAAqB;IACpC,OAAO,CAAC,OAAO,CAAsB;IACrC,OAAO,CAAC,YAAY,CAAqB;IACzC,OAAO,CAAC,OAAO,CAAK;IACpB,OAAO,CAAC,QAAQ,CAAS;IACzB,OAAO,CAAC,IAAI,CAAc;IAC1B,oEAAoE;IACpE,OAAO,CAAC,UAAU,CAAK;IACvB,OAAO,CAAC,SAAS,CAA4C;IAC7D,OAAO,CAAC,OAAO,CAAK;IACpB,gEAAgE;IAChE,OAAO,CAAC,UAAU,CAAK;IACvB,kGAAkG;IAClG,OAAO,CAAC,WAAW,CAAK;IACxB,gGAAgG;IAChG,OAAO,CAAC,YAAY,CAAS;IAC7B,OAAO,CAAC,IAAI,CAAkB;IAC9B,8FAA8F;IAC9F,OAAO,CAAC,aAAa,CAAS;IAC9B,kEAAkE;IAClE,OAAO,CAAC,aAAa,CAAS;gBAED,OAAO,EAAE,UAAU;IAIhD;;;;;OAKG;IACG,IAAI,CAAC,SAAS,SAAc,GAAG,OAAO,CAAC,MAAM,CAAC;IAuCpD;;OAEG;IACH,OAAO,CAAC,KAAK,EAAE,MAAM,EAAE,EAAE,EAAE,MAAM,EAAE,KAAK,UAAQ,GAAG,IAAI;IAIvD;;;;OAIG;IACH,MAAM,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,IAAI;IAShC;;;OAGG;IACH,SAAS,IAAI,IAAI;IAKjB,kFAAkF;IAClF,IAAI,IAAI,IAAI;IAgBZ,OAAO,CAAC,IAAI;IAcZ,OAAO,CAAC,OAAO;IAqBf;;;OAGG;IACH,OAAO,CAAC,OAAO;IAcf;;;OAGG;IACH,OAAO,CAAC,SAAS;IAkBjB,sGAAsG;IACtG,OAAO,CAAC,YAAY;IAkBpB,OAAO,CAAC,GAAG;IAIX,OAAO,CAAC,UAAU;IAIlB,6DAA6D;IAC7D,OAAO,CAAC,UAAU;IAqClB;;;OAGG;IACH,OAAO,CAAC,MAAM;IAYd,OAAO,CAAC,IAAI;CAKb"}

package/dist/ear.js

@@ -0,0 +1,387 @@
+import { silent } from './log.js';
+import { EchoGate } from './barge-in.js';
+import { TURN_COMPLETE } from './turn.js';
+import { FRAME_SAMPLES, VoiceDetector } from './vad.js';
+export const MAX_WAIT_MS = 60_000;
+/**
+ * What the AI said may come back through the microphone (echo) up to this long after it finished saying it.
+ */
+export const ECHO_WINDOW_MS = 10_000;
+/** A sentence is echo if at least this share of its words came from what the AI said… */
+export const ECHO_OVERLAP = 0.7;
+/** …and it has at least this many words (a user's "yes" or "ok" is never echo). */
+export const ECHO_MIN_WORDS = 3;
+/**
+ * If the AI does not call listen again within this time, the conversation is over and the microphone closes.
+ */
+export const IDLE_CLOSE_MS = 30_000;
+/** A "sentence" with less voice energy than this is a knock or a noise and is not transcribed. */
+export const MIN_VOICED_MS = 300;
+/**
+ * With something already understood, listen waits at most this long for the speaker to stop (it may be a TV).
+ */
+export const SETTLE_MS = 3_000;
+/**
+ * At every 0.3 s pause (10 blocks) Smart Turn is asked whether the user has finished. If not, the 1.5 s limit
+ * applies.
+ */
+export const PAUSE_FRAMES = 10;
+/**
+ * Smart Turn only closes a sentence early when it has at least this much voice, so it does not split a short
+ * question in pieces.
+ */
+export const MIN_TURN_VOICED_MS = 1_000;
+/**
+ * When the user cuts the AI off, their sentence starts this much earlier (17 blocks, 0.5 s); what came before
+ * was the AI.
+ */
+const BARGE_KEEP_FRAMES = 17;
+/**
+ * A sentence that started while the AI was talking and is not its echo only cuts the AI off with at least
+ * this much voice. Echo leftovers are chunks with little voice (330 ms in 3 s) that Whisper sometimes
+ * transcribes as other words; a user saying "stop, stop" had 750 ms.
+ */
+export const MIN_INTERRUPT_VOICED_MS = 500;
+/**
+ * …and, under 3 words, it must be almost all voice. An echo leftover is a few words in a long chunk with
+ * little voice (510 ms of voice in 3.4 s, 15 %); a single "stop" from the user is almost all voice.
+ */
+export const MIN_INTERRUPT_VOICE_RATIO = 0.3;
+/**
+ * Whether it is a leftover of the AI's echo, even if Whisper transcribed it with words the AI did not say.
+ */
+function echoLeftover(text, voicedMs, audioMs) {
+    const words = (text.match(/\p{L}+/gu) ?? []).length;
+    return voicedMs < MIN_INTERRUPT_VOICED_MS || (words < 3 && voicedMs / audioMs < MIN_INTERRUPT_VOICE_RATIO);
+}
+/**
+ * What listen returns when the user cut the AI off, so the AI knows the user did not hear everything it said.
+ */
+export const INTERRUPTED = '(the user interrupted you while you were speaking)';
+const rms = (frame) => {
+    let sum = 0;
+    for (const sample of frame)
+        sum += sample * sample;
+    return Math.sqrt(sum / frame.length);
+};
+/** Whisper sometimes makes up an endlessly repeated word over noise or music ("yes, yes, yes…"). */
+function isRepetition(text) {
+    const all = text.toLowerCase().match(/\p{L}+/gu) ?? [];
+    return all.length >= 8 && new Set(all).size / all.length < 0.25;
+}
+/** Words without accents or capitals, to compare what was heard with what was said. */
+function words(text) {
+    const plain = text.toLowerCase().normalize('NFD').replace(/[̀-ͯ]/g, '');
+    return new Set(plain.match(/[a-z0-9ñ]+/g) ?? []);
+}
+/**
+ * Continuous listening. The microphone stays open from the first listen, also while the AI talks, so the user
+ * can answer or interrupt at once. Each long pause closes a sentence, which is transcribed and stored; listen
+ * returns it whole. Whatever matches what the AI just said is its own echo from the speakers and is
+ * discarded.
+ */
+export class Ear {
+    options;
+    mic;
+    vad;
+    leftover = Buffer.alloc(0);
+    /**
+     * Understood text not collected by listen yet. `fromUser`: it started while the AI was quiet, so it is not
+     * an echo leftover.
+     */
+    heard = [];
+    failure;
+    waiters = [];
+    transcribing = Promise.resolve();
+    pending = 0;
+    speaking = false;
+    said = [];
+    /** Bumped on close, so transcriptions in progress are discarded. */
+    generation = 0;
+    idleTimer;
+    heardAt = 0;
+    /** Bumped on every interrupt, so the waiting listen returns. */
+    interrupts = 0;
+    /** Bumped on every closed sentence, so a Smart Turn answer for a previous sentence is ignored. */
+    utteranceId = 0;
+    /** Smart Turn is deciding: it is not asked again, since each pass delays Silero and Whisper. */
+    checkingTurn = false;
+    gate = new EchoGate();
+    /** The current sentence started while the AI was talking, so it is almost always its echo. */
+    startedWithAI = false;
+    /** The user cut the AI off; the AI is told on the next listen. */
+    interruptedAI = false;
+    constructor(options) {
+        this.options = options;
+        this.vad = new VoiceDetector(options.vad);
+    }
+    /**
+     * Everything the user said since the last call, whole: waits for them to stop talking and for the rest to
+     * be transcribed. It never gets stuck with endless background voice (a TV, a video): with something already
+     * understood, it returns at most SETTLE_MS later, and after maxWaitMs it returns whatever there is (or
+     * nothing). Anything still playing is not lost: it goes to the next listen.
+     */
+    async next(maxWaitMs = MAX_WAIT_MS) {
+        clearTimeout(this.idleTimer);
+        this.open();
+        const interrupts = this.interrupts;
+        let timedOut = false;
+        const timer = setTimeout(() => {
+            timedOut = true;
+            this.wake();
+        }, maxWaitMs);
+        let settleTimer;
+        const busy = () => this.speaking || this.pending > 0;
+        while (!this.failure && !timedOut && interrupts === this.interrupts) {
+            if (this.heard.length) {
+                const left = this.heardAt + SETTLE_MS - Date.now();
+                if (!busy() || left <= 0)
+                    break;
+                settleTimer ??= setTimeout(() => this.wake(), left);
+            }
+            await new Promise((resolve) => this.waiters.push(resolve));
+        }
+        clearTimeout(timer);
+        clearTimeout(settleTimer);
+        // Another tool interrupted it: it returns without consuming what was heard, which is kept for the next
+        // listen.
+        if (interrupts !== this.interrupts)
+            return '';
+        // If the AI does not listen again soon, the conversation is over and the microphone does not stay open.
+        this.idleTimer = setTimeout(() => this.stop(), IDLE_CLOSE_MS);
+        if (this.failure) {
+            const reason = this.failure;
+            this.failure = undefined;
+            throw new Error(reason);
+        }
+        let text = this.heard.map((h) => h.text).join(' ');
+        this.heard = [];
+        // The AI is told it was cut off once, even if what cut it off never became a sentence.
+        if (text && this.interruptedAI)
+            text = `${INTERRUPTED} ${text}`;
+        this.interruptedAI = false;
+        return text;
+    }
+    /**
+     * A 30 ms chunk about to play from the speaker (its level and when it plays). `first` starts a sentence.
+     */
+    playing(level, at, first = false) {
+        this.gate.play(level, at, first);
+    }
+    /**
+     * The AI is about to say this text. Returns the function to call when it finishes: it closes the current
+     * sentence, which is its echo from the speakers, so it does not merge with what the user says next (if they
+     * merged, the echo filter would also discard the user's words).
+     */
+    saying(text) {
+        const said = { words: words(text), until: Infinity };
+        this.said.push(said);
+        return () => {
+            said.until = Math.max(Number.isFinite(said.until) ? said.until : 0, Date.now() + ECHO_WINDOW_MS);
+            if (!said.cut && this.vad.current === 'speaking')
+                this.endUtterance();
+        };
+    }
+    /**
+     * Another tool arrives (speak, stop or another listen): the waiting listen returns now, without consuming
+     * what was heard. This way the AI can speak even if an earlier listen is still waiting.
+     */
+    interrupt() {
+        this.interrupts++;
+        this.wake();
+    }
+    /** Ends the conversation: closes the microphone and clears everything pending. */
+    stop() {
+        clearTimeout(this.idleTimer);
+        this.mic?.stop();
+        this.mic = undefined;
+        this.generation++;
+        this.utteranceId++;
+        this.gate = new EchoGate();
+        this.interruptedAI = false;
+        this.heard = [];
+        this.said = [];
+        this.speaking = false;
+        this.leftover = Buffer.alloc(0);
+        this.vad = new VoiceDetector(this.options.vad);
+        this.wake();
+    }
+    open() {
+        if (this.mic)
+            return;
+        const mic = this.options.openMic();
+        this.mic = mic;
+        mic.onData((chunk) => this.onAudio(chunk));
+        // If the microphone closes by itself, the next listen opens it again.
+        mic.onExit((reason) => {
+            if (this.mic !== mic)
+                return;
+            this.mic = undefined;
+            this.failure = `Cannot use the microphone: ${reason}`;
+            this.wake();
+        });
+    }
+    onAudio(chunk) {
+        if (!this.mic)
+            return;
+        const bytes = Buffer.concat([this.leftover, chunk]);
+        const frameBytes = FRAME_SAMPLES * 2;
+        let offset = 0;
+        for (; offset + frameBytes <= bytes.length; offset += frameBytes) {
+            const frame = new Float32Array(FRAME_SAMPLES);
+            for (let i = 0; i < FRAME_SAMPLES; i++)
+                frame[i] = bytes.readInt16LE(offset + i * 2) / 32_768;
+            if (this.aiSpeaking() && this.gate.mic(rms(frame), Date.now(), this.vad.threshold))
+                this.bargeIn();
+            const was = this.speaking;
+            // No noise is learned while the AI talks: after a minute of reading, its echo passed for noise and the user's
+            // "stop" no longer crossed the threshold.
+            const state = this.vad.push(frame, !this.aiSpeaking());
+            this.speaking = state === 'speaking';
+            if (this.speaking && !was)
+                this.startedWithAI = this.aiSpeaking();
+            if (state === 'done')
+                this.endUtterance();
+            else if (state === 'speaking' && this.vad.quietFrames === PAUSE_FRAMES)
+                this.checkTurn();
+        }
+        this.leftover = bytes.subarray(offset);
+    }
+    /**
+     * The user talks over the AI: it goes quiet, and the user's sentence starts here (what came before was the
+     * AI's voice, which the echo filter would discard along with the user's words).
+     */
+    bargeIn() {
+        this.log('the user is talking: cutting the AI off');
+        for (const said of this.said) {
+            if (said.until !== Infinity)
+                continue;
+            said.cut = true;
+            said.until = Date.now() + ECHO_WINDOW_MS;
+        }
+        this.interruptedAI = true;
+        // From here on, the current sentence belongs to the user.
+        this.startedWithAI = false;
+        if (this.vad.current === 'speaking')
+            this.vad.keepLast(BARGE_KEEP_FRAMES);
+        this.options.onBargeIn?.();
+    }
+    /**
+     * 0.3 s pause: Smart Turn tells whether the user has finished. If so and they are still quiet, the sentence
+     * closes now instead of after 1.5 s. If they started talking again meanwhile, the answer is ignored.
+     */
+    checkTurn() {
+        const turnCheck = this.options.turnCheck;
+        // While the AI talks, what is heard is almost always its echo: the sentence closes when it stops talking.
+        if (!turnCheck || this.checkingTurn || this.vad.voicedMs < MIN_TURN_VOICED_MS || this.aiSpeaking())
+            return;
+        this.checkingTurn = true;
+        const id = this.utteranceId;
+        const started = Date.now();
+        turnCheck.complete(this.vad.utterance()).then((probability) => {
+            const still = id === this.utteranceId && this.vad.current === 'speaking' && this.vad.quietFrames >= PAUSE_FRAMES;
+            const verdict = probability > TURN_COMPLETE ? 'finished' : 'just a pause';
+            this.log(`Smart Turn: ${verdict} (${probability.toFixed(2)}, ${Date.now() - started} ms)${still ? '' : ' · outdated'}`);
+            if (still && probability > TURN_COMPLETE)
+                this.endUtterance();
+        }, (error) => this.log(`Smart Turn failed: ${error.message}`)).finally(() => (this.checkingTurn = false));
+    }
+    /** Closes the current sentence and prepares the next one; the detector keeps the room noise level. */
+    endUtterance() {
+        const audio = this.vad.utterance();
+        const voiced = this.vad.voicedMs;
+        const fromUser = !this.startedWithAI;
+        this.vad.next();
+        this.utteranceId++;
+        this.speaking = false;
+        const ms = Math.round((audio.length / 16_000) * 1000);
+        if (voiced >= MIN_VOICED_MS) {
+            this.log(`sentence closed: ${ms} ms of audio, ${voiced} ms of voice${this.aiSpeaking() ? ' (while the AI was talking)' : ''}`);
+            if (this.waiters.length && !this.aiSpeaking())
+                this.options.onPhrase?.();
+            this.transcribe(audio, fromUser, voiced, ms);
+        }
+        else {
+            this.log(`stray noise dropped: ${ms} ms, ${voiced} ms of voice`);
+            this.wake();
+        }
+    }
+    log(message) {
+        (this.options.log ?? silent)(`[ear] ${message}`);
+    }
+    aiSpeaking() {
+        return this.said.some((said) => said.until === Infinity);
+    }
+    /** Sentences are transcribed in order, one after another. */
+    transcribe(audio, fromUser, voiced = Infinity, audioMs = Infinity) {
+        const generation = this.generation;
+        this.pending++;
+        this.transcribing = this.transcribing.then(async () => {
+            try {
+                if (this.options.speechCheck && !(await this.options.speechCheck.hasSpeech(audio))) {
+                    this.log('Silero: not human voice → not transcribed');
+                    return;
+                }
+                const text = (await this.options.transcriber.transcribe(audio)).trim();
+                if (generation !== this.generation)
+                    return this.log(`dropped (the microphone closed meanwhile): «${text}»`);
+                if (!text)
+                    return this.log('Whisper returned no text');
+                // What started while the AI was talking is almost always its echo: it is dropped if nearly all its words are
+                // the AI's, however short.
+                if (this.isEcho(text, fromUser ? ECHO_MIN_WORDS : 1))
+                    return this.log(`AI echo dropped: «${text}»`);
+                if (isRepetition(text))
+                    return this.log(`endless repetition dropped: «${text}»`);
+                if (!fromUser && echoLeftover(text, voiced, audioMs))
+                    return this.log(`AI echo leftover dropped (${voiced} ms of voice in ${audioMs} ms): «${text}»`);
+                this.log(`heard: «${text}»`);
+                if (!fromUser) {
+                    // It started while the AI was talking but is not its echo: the user is talking over it (their "stop" was not
+                    // in what the AI said). This is the second way to cut the AI off, in case the microphone level check did not
+                    // fire.
+                    this.log('the user talked over the AI (by what they said)');
+                    if (this.aiSpeaking())
+                        this.bargeIn();
+                    else
+                        this.interruptedAI = true;
+                }
+                if (!this.heard.length)
+                    this.heardAt = Date.now();
+                this.heard.push({ text, fromUser: true });
+            }
+            catch (error) {
+                this.failure = `Could not transcribe the sentence: ${error instanceof Error ? error.message : String(error)}`;
+            }
+            finally {
+                this.pending--;
+                this.wake();
+            }
+        });
+    }
+    /**
+     * With fewer than `minWords` words it is never echo (a user's "yes"); 1 for what started while the AI was
+     * talking.
+     */
+    isEcho(text, minWords = ECHO_MIN_WORDS) {
+        const now = Date.now();
+        this.said = this.said.filter((said) => said.until > now);
+        const heard = words(text);
+        if (heard.size < minWords)
+            return false;
+        return this.said.some((said) => {
+            let shared = 0;
+            for (const word of heard)
+                if (said.words.has(word))
+                    shared++;
+            return shared / heard.size >= ECHO_OVERLAP;
+        });
+    }
+    wake() {
+        const waiters = this.waiters;
+        this.waiters = [];
+        for (const waiter of waiters)
+            waiter();
+    }
+}
+//# sourceMappingURL=ear.js.map

package/dist/ear.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"ear.js","sourceRoot":"","sources":["../src/ear.ts"],"names":[],"mappings":"AAAA,OAAO,EAAY,MAAM,EAAE,MAAM,UAAU,CAAC;AAI5C,OAAO,EAAE,QAAQ,EAAE,MAAM,eAAe,CAAC;AACzC,OAAO,EAAE,aAAa,EAAkB,MAAM,WAAW,CAAC;AAC1D,OAAO,EAAE,aAAa,EAAmB,aAAa,EAAE,MAAM,UAAU,CAAC;AAEzE,MAAM,CAAC,MAAM,WAAW,GAAG,MAAM,CAAC;AAClC;;GAEG;AACH,MAAM,CAAC,MAAM,cAAc,GAAG,MAAM,CAAC;AACrC,yFAAyF;AACzF,MAAM,CAAC,MAAM,YAAY,GAAG,GAAG,CAAC;AAChC,mFAAmF;AACnF,MAAM,CAAC,MAAM,cAAc,GAAG,CAAC,CAAC;AAChC;;GAEG;AACH,MAAM,CAAC,MAAM,aAAa,GAAG,MAAM,CAAC;AACpC,kGAAkG;AAClG,MAAM,CAAC,MAAM,aAAa,GAAG,GAAG,CAAC;AACjC;;GAEG;AACH,MAAM,CAAC,MAAM,SAAS,GAAG,KAAK,CAAC;AAC/B;;;GAGG;AACH,MAAM,CAAC,MAAM,YAAY,GAAG,EAAE,CAAC;AAC/B;;;GAGG;AACH,MAAM,CAAC,MAAM,kBAAkB,GAAG,KAAK,CAAC;AACxC;;;GAGG;AACH,MAAM,iBAAiB,GAAG,EAAE,CAAC;AAC7B;;;;GAIG;AACH,MAAM,CAAC,MAAM,uBAAuB,GAAG,GAAG,CAAC;AAC3C;;;GAGG;AACH,MAAM,CAAC,MAAM,yBAAyB,GAAG,GAAG,CAAC;AAE7C;;GAEG;AACH,SAAS,YAAY,CAAC,IAAY,EAAE,QAAgB,EAAE,OAAe;IACnE,MAAM,KAAK,GAAG,CAAC,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,IAAI,EAAE,CAAC,CAAC,MAAM,CAAC;IACpD,OAAO,QAAQ,GAAG,uBAAuB,IAAI,CAAC,KAAK,GAAG,CAAC,IAAI,QAAQ,GAAG,OAAO,GAAG,yBAAyB,CAAC,CAAC;AAC7G,CAAC;AACD;;GAEG;AACH,MAAM,CAAC,MAAM,WAAW,GAAG,oDAAoD,CAAC;AAEhF,MAAM,GAAG,GAAG,CAAC,KAAmB,EAAE,EAAE;IAClC,IAAI,GAAG,GAAG,CAAC,CAAC;IACZ,KAAK,MAAM,MAAM,IAAI,KAAK;QAAE,GAAG,IAAI,MAAM,GAAG,MAAM,CAAC;IACnD,OAAO,IAAI,CAAC,IAAI,CAAC,GAAG,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC;AACvC,CAAC,CAAC;AAEF,oGAAoG;AACpG,SAAS,YAAY,CAAC,IAAY;IAChC,MAAM,GAAG,GAAG,IAAI,CAAC,WAAW,EAAE,CAAC,KAAK,CAAC,UAAU,CAAC,IAAI,EAAE,CAAC;IACvD,OAAO,GAAG,CAAC,MAAM,IAAI,CAAC,IAAI,IAAI,GAAG,CAAC,GAAG,CAAC,CAAC,IAAI,GAAG,GAAG,CAAC,MAAM,GAAG,IAAI,CAAC;AAClE,CAAC;AA+BD,uFAAuF;AACvF,SAAS,KAAK,CAAC,IAAY;IACzB,MAAM,KAAK,GAAG,IAAI,CAAC,WAAW,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,OAAO,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC;IACxE,OAAO,IAAI,GAAG,CAAC,KAAK,CAAC,KAAK,CAAC,aAAa,CAAC,IAAI,EAAE,CAAC,CAAC;AACnD,CAAC;AAED;;;;;GAKG;AACH,MAAM,OAAO,GAAG;IA+Be;IA9BrB,GAAG,CAAkB;IACrB,GAAG,CAAgB;IACnB,QAAQ,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IACnC;;;OAGG;IACK,KAAK,GAA0C,EAAE,CAAC;IAClD,OAAO,CAAqB;IAC5B,OAAO,GAAmB,EAAE,CAAC;IAC7B,YAAY,GAAG,OAAO,CAAC,OAAO,EAAE,CAAC;IACjC,OAAO,GAAG,CAAC,CAAC;IACZ,QAAQ,GAAG,KAAK,CAAC;IACjB,IAAI,GAAW,EAAE,CAAC;IAC1B,oEAAoE;IAC5D,UAAU,GAAG,CAAC,CAAC;IACf,SAAS,CAA4C;IACrD,OAAO,GAAG,CAAC,CAAC;IACpB,gEAAgE;IACxD,UAAU,GAAG,CAAC,CAAC;IACvB,kGAAkG;IAC1F,WAAW,GAAG,CAAC,CAAC;IACxB,gGAAgG;IACxF,YAAY,GAAG,KAAK,CAAC;IACrB,IAAI,GAAG,IAAI,QAAQ,EAAE,CAAC;IAC9B,8FAA8F;IACtF,aAAa,GAAG,KAAK,CAAC;IAC9B,kEAAkE;IAC1D,aAAa,GAAG,KAAK,CAAC;IAE9B,YAA6B,OAAmB;QAAnB,YAAO,GAAP,OAAO,CAAY;QAC9C,IAAI,CAAC,GAAG,GAAG,IAAI,aAAa,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;IAC5C,CAAC;IAED;;;;;OAKG;IACH,KAAK,CAAC,IAAI,CAAC,SAAS,GAAG,WAAW;QAChC,YAAY,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;QAC7B,IAAI,CAAC,IAAI,EAAE,CAAC;QACZ,MAAM,UAAU,GAAG,IAAI,CAAC,UAAU,CAAC;QACnC,IAAI,QAAQ,GAAG,KAAK,CAAC;QACrB,MAAM,KAAK,GAAG,UAAU,CAAC,GAAG,EAAE;YAC5B,QAAQ,GAAG,IAAI,CAAC;YAChB,IAAI,CAAC,IAAI,EAAE,CAAC;QACd,CAAC,EAAE,SAAS,CAAC,CAAC;QACd,IAAI,WAAsD,CAAC;QAC3D,MAAM,IAAI,GAAG,GAAG,EAAE,CAAC,IAAI,CAAC,QAAQ,IAAI,IAAI,CAAC,OAAO,GAAG,CAAC,CAAC;QACrD,OAAO,CAAC,IAAI,CAAC,OAAO,IAAI,CAAC,QAAQ,IAAI,UAAU,KAAK,IAAI,CAAC,UAAU,EAAE,CAAC;YACpE,IAAI,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC;gBACtB,MAAM,IAAI,GAAG,IAAI,CAAC,OAAO,GAAG,SAAS,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;gBACnD,IAAI,CAAC,IAAI,EAAE,IAAI,IAAI,IAAI,CAAC;oBAAE,MAAM;gBAChC,WAAW,KAAK,UAAU,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,IAAI,EAAE,EAAE,IAAI,CAAC,CAAC;YACtD,CAAC;YACD,MAAM,IAAI,OAAO,CAAO,CAAC,OAAO,EAAE,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC;QACnE,CAAC;QACD,YAAY,CAAC,KAAK,CAAC,CAAC;QACpB,YAAY,CAAC,WAAW,CAAC,CAAC;QAC9B,uGAAuG;QACvG,UAAU;QACN,IAAI,UAAU,KAAK,IAAI,CAAC,UAAU;YAAE,OAAO,EAAE,CAAC;QAClD,wGAAwG;QACpG,IAAI,CAAC,SAAS,GAAG,UAAU,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,IAAI,EAAE,EAAE,aAAa,CAAC,CAAC;QAC9D,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YACjB,MAAM,MAAM,GAAG,IAAI,CAAC,OAAO,CAAC;YAC5B,IAAI,CAAC,OAAO,GAAG,SAAS,CAAC;YACzB,MAAM,IAAI,KAAK,CAAC,MAAM,CAAC,CAAC;QAC1B,CAAC;QACD,IAAI,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;QACnD,IAAI,CAAC,KAAK,GAAG,EAAE,CAAC;QACpB,uFAAuF;QACnF,IAAI,IAAI,IAAI,IAAI,CAAC,aAAa;YAAE,IAAI,GAAG,GAAG,WAAW,IAAI,IAAI,EAAE,CAAC;QAChE,IAAI,CAAC,aAAa,GAAG,KAAK,CAAC;QAC3B,OAAO,IAAI,CAAC;IACd,CAAC;IAED;;OAEG;IACH,OAAO,CAAC,KAAa,EAAE,EAAU,EAAE,KAAK,GAAG,KAAK;QAC9C,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,KAAK,EAAE,EAAE,EAAE,KAAK,CAAC,CAAC;IACnC,CAAC;IAED;;;;OAIG;IACH,MAAM,CAAC,IAAY;QACjB,MAAM,IAAI,GAAS,EAAE,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,QAAQ,EAAE,CAAC;QAC3D,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACrB,OAAO,GAAG,EAAE;YACV,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,EAAE,GAAG,cAAc,CAAC,CAAC;YACjG,IAAI,CAAC,IAAI,CAAC,GAAG,IAAI,IAAI,CAAC,GAAG,CAAC,OAAO,KAAK,UAAU;gBAAE,IAAI,CAAC,YAAY,EAAE,CAAC;QACxE,CAAC,CAAC;IACJ,CAAC;IAED;;;OAGG;IACH,SAAS;QACP,IAAI,CAAC,UAAU,EAAE,CAAC;QAClB,IAAI,CAAC,IAAI,EAAE,CAAC;IACd,CAAC;IAED,kFAAkF;IAClF,IAAI;QACF,YAAY,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;QAC7B,IAAI,CAAC,GAAG,EAAE,IAAI,EAAE,CAAC;QACjB,IAAI,CAAC,GAAG,GAAG,SAAS,CAAC;QACrB,IAAI,CAAC,UAAU,EAAE,CAAC;QAClB,IAAI,CAAC,WAAW,EAAE,CAAC;QACnB,IAAI,CAAC,IAAI,GAAG,IAAI,QAAQ,EAAE,CAAC;QAC3B,IAAI,CAAC,aAAa,GAAG,KAAK,CAAC;QAC3B,IAAI,CAAC,KAAK,GAAG,EAAE,CAAC;QAChB,IAAI,CAAC,IAAI,GAAG,EAAE,CAAC;QACf,IAAI,CAAC,QAAQ,GAAG,KAAK,CAAC;QACtB,IAAI,CAAC,QAAQ,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QAChC,IAAI,CAAC,GAAG,GAAG,IAAI,aAAa,CAAC,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;QAC/C,IAAI,CAAC,IAAI,EAAE,CAAC;IACd,CAAC;IAEO,IAAI;QACV,IAAI,IAAI,CAAC,GAAG;YAAE,OAAO;QACrB,MAAM,GAAG,GAAG,IAAI,CAAC,OAAO,CAAC,OAAO,EAAE,CAAC;QACnC,IAAI,CAAC,GAAG,GAAG,GAAG,CAAC;QACf,GAAG,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC;QAC/C,sEAAsE;QAClE,GAAG,CAAC,MAAM,CAAC,CAAC,MAAM,EAAE,EAAE;YACpB,IAAI,IAAI,CAAC,GAAG,KAAK,GAAG;gBAAE,OAAO;YAC7B,IAAI,CAAC,GAAG,GAAG,SAAS,CAAC;YACrB,IAAI,CAAC,OAAO,GAAG,8BAA8B,MAAM,EAAE,CAAC;YACtD,IAAI,CAAC,IAAI,EAAE,CAAC;QACd,CAAC,CAAC,CAAC;IACL,CAAC;IAEO,OAAO,CAAC,KAAa;QAC3B,IAAI,CAAC,IAAI,CAAC,GAAG;YAAE,OAAO;QACtB,MAAM,KAAK,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,IAAI,CAAC,QAAQ,EAAE,KAAK,CAAC,CAAC,CAAC;QACpD,MAAM,UAAU,GAAG,aAAa,GAAG,CAAC,CAAC;QACrC,IAAI,MAAM,GAAG,CAAC,CAAC;QACf,OAAO,MAAM,GAAG,UAAU,IAAI,KAAK,CAAC,MAAM,EAAE,MAAM,IAAI,UAAU,EAAE,CAAC;YACjE,MAAM,KAAK,GAAG,IAAI,YAAY,CAAC,aAAa,CAAC,CAAC;YAC9C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,aAAa,EAAE,CAAC,EAAE;gBAAE,KAAK,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC,WAAW,CAAC,MAAM,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,MAAM,CAAC;YAC9F,IAAI,IAAI,CAAC,UAAU,EAAE,IAAI,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,IAAI,CAAC,GAAG,EAAE,EAAE,IAAI,CAAC,GAAG,CAAC,SAAS,CAAC;gBAAE,IAAI,CAAC,OAAO,EAAE,CAAC;YACnG,MAAM,GAAG,GAAG,IAAI,CAAC,QAAQ,CAAC;YAChC,8GAA8G;YAC9G,0CAA0C;YACpC,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,KAAK,EAAE,CAAC,IAAI,CAAC,UAAU,EAAE,CAAC,CAAC;YACvD,IAAI,CAAC,QAAQ,GAAG,KAAK,KAAK,UAAU,CAAC;YACrC,IAAI,IAAI,CAAC,QAAQ,IAAI,CAAC,GAAG;gBAAE,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC,UAAU,EAAE,CAAC;YAClE,IAAI,KAAK,KAAK,MAAM;gBAAE,IAAI,CAAC,YAAY,EAAE,CAAC;iBACrC,IAAI,KAAK,KAAK,UAAU,IAAI,IAAI,CAAC,GAAG,CAAC,WAAW,KAAK,YAAY;gBAAE,IAAI,CAAC,SAAS,EAAE,CAAC;QAC3F,CAAC;QACD,IAAI,CAAC,QAAQ,GAAG,KAAK,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC;IACzC,CAAC;IAED;;;OAGG;IACK,OAAO;QACb,IAAI,CAAC,GAAG,CAAC,yCAAyC,CAAC,CAAC;QACpD,KAAK,MAAM,IAAI,IAAI,IAAI,CAAC,IAAI,EAAE,CAAC;YAC7B,IAAI,IAAI,CAAC,KAAK,KAAK,QAAQ;gBAAE,SAAS;YACtC,IAAI,CAAC,GAAG,GAAG,IAAI,CAAC;YAChB,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,cAAc,CAAC;QAC3C,CAAC;QACD,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC;QAC9B,0DAA0D;QACtD,IAAI,CAAC,aAAa,GAAG,KAAK,CAAC;QAC3B,IAAI,IAAI,CAAC,GAAG,CAAC,OAAO,KAAK,UAAU;YAAE,IAAI,CAAC,GAAG,CAAC,QAAQ,CAAC,iBAAiB,CAAC,CAAC;QAC1E,IAAI,CAAC,OAAO,CAAC,SAAS,EAAE,EAAE,CAAC;IAC7B,CAAC;IAED;;;OAGG;IACK,SAAS;QACf,MAAM,SAAS,GAAG,IAAI,CAAC,OAAO,CAAC,SAAS,CAAC;QAC7C,0GAA0G;QACtG,IAAI,CAAC,SAAS,IAAI,IAAI,CAAC,YAAY,IAAI,IAAI,CAAC,GAAG,CAAC,QAAQ,GAAG,kBAAkB,IAAI,IAAI,CAAC,UAAU,EAAE;YAAE,OAAO;QAC3G,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC;QACzB,MAAM,EAAE,GAAG,IAAI,CAAC,WAAW,CAAC;QAC5B,MAAM,OAAO,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QAC3B,SAAS,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,CAAC,CAAC,IAAI,CAC3C,CAAC,WAAW,EAAE,EAAE;YACd,MAAM,KAAK,GAAG,EAAE,KAAK,IAAI,CAAC,WAAW,IAAI,IAAI,CAAC,GAAG,CAAC,OAAO,KAAK,UAAU,IAAI,IAAI,CAAC,GAAG,CAAC,WAAW,IAAI,YAAY,CAAC;YACjH,MAAM,OAAO,GAAG,WAAW,GAAG,aAAa,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,cAAc,CAAC;YAC1E,IAAI,CAAC,GAAG,CAAC,eAAe,OAAO,KAAK,WAAW,CAAC,OAAO,CAAC,CAAC,CAAC,KAAK,IAAI,CAAC,GAAG,EAAE,GAAG,OAAO,OAAO,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,aAAa,EAAE,CAAC,CAAC;YACxH,IAAI,KAAK,IAAI,WAAW,GAAG,aAAa;gBAAE,IAAI,CAAC,YAAY,EAAE,CAAC;QAChE,CAAC,EACD,CAAC,KAAY,EAAE,EAAE,CAAC,IAAI,CAAC,GAAG,CAAC,sBAAsB,KAAK,CAAC,OAAO,EAAE,CAAC,CAClE,CAAC,OAAO,CAAC,GAAG,EAAE,CAAC,CAAC,IAAI,CAAC,YAAY,GAAG,KAAK,CAAC,CAAC,CAAC;IAC/C,CAAC;IAED,sGAAsG;IAC9F,YAAY;QAClB,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,CAAC;QACnC,MAAM,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,QAAQ,CAAC;QACjC,MAAM,QAAQ,GAAG,CAAC,IAAI,CAAC,aAAa,CAAC;QACrC,IAAI,CAAC,GAAG,CAAC,IAAI,EAAE,CAAC;QAChB,IAAI,CAAC,WAAW,EAAE,CAAC;QACnB,IAAI,CAAC,QAAQ,GAAG,KAAK,CAAC;QACtB,MAAM,EAAE,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,MAAM,GAAG,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC;QACtD,IAAI,MAAM,IAAI,aAAa,EAAE,CAAC;YAC5B,IAAI,CAAC,GAAG,CAAC,oBAAoB,EAAE,iBAAiB,MAAM,eAAe,IAAI,CAAC,UAAU,EAAE,CAAC,CAAC,CAAC,6BAA6B,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;YAC/H,IAAI,IAAI,CAAC,OAAO,CAAC,MAAM,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE;gBAAE,IAAI,CAAC,OAAO,CAAC,QAAQ,EAAE,EAAE,CAAC;YACzE,IAAI,CAAC,UAAU,CAAC,KAAK,EAAE,QAAQ,EAAE,MAAM,EAAE,EAAE,CAAC,CAAC;QAC/C,CAAC;aAAM,CAAC;YACN,IAAI,CAAC,GAAG,CAAC,wBAAwB,EAAE,QAAQ,MAAM,cAAc,CAAC,CAAC;YACjE,IAAI,CAAC,IAAI,EAAE,CAAC;QACd,CAAC;IACH,CAAC;IAEO,GAAG,CAAC,OAAe;QACzB,CAAC,IAAI,CAAC,OAAO,CAAC,GAAG,IAAI,MAAM,CAAC,CAAC,SAAS,OAAO,EAAE,CAAC,CAAC;IACnD,CAAC;IAEO,UAAU;QAChB,OAAO,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,KAAK,KAAK,QAAQ,CAAC,CAAC;IAC3D,CAAC;IAED,6DAA6D;IACrD,UAAU,CAAC,KAAmB,EAAE,QAAiB,EAAE,MAAM,GAAG,QAAQ,EAAE,OAAO,GAAG,QAAQ;QAC9F,MAAM,UAAU,GAAG,IAAI,CAAC,UAAU,CAAC;QACnC,IAAI,CAAC,OAAO,EAAE,CAAC;QACf,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC,YAAY,CAAC,IAAI,CAAC,KAAK,IAAI,EAAE;YACpD,IAAI,CAAC;gBACH,IAAI,IAAI,CAAC,OAAO,CAAC,WAAW,IAAI,CAAC,CAAC,MAAM,IAAI,CAAC,OAAO,CAAC,WAAW,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC;oBACnF,IAAI,CAAC,GAAG,CAAC,2CAA2C,CAAC,CAAC;oBACtD,OAAO;gBACT,CAAC;gBACD,MAAM,IAAI,GAAG,CAAC,MAAM,IAAI,CAAC,OAAO,CAAC,WAAW,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC;gBACvE,IAAI,UAAU,KAAK,IAAI,CAAC,UAAU;oBAAE,OAAO,IAAI,CAAC,GAAG,CAAC,+CAA+C,IAAI,GAAG,CAAC,CAAC;gBAC5G,IAAI,CAAC,IAAI;oBAAE,OAAO,IAAI,CAAC,GAAG,CAAC,0BAA0B,CAAC,CAAC;gBAC/D,6GAA6G;gBAC7G,2BAA2B;gBACnB,IAAI,IAAI,CAAC,MAAM,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC,CAAC,cAAc,CAAC,CAAC,CAAC,CAAC,CAAC;oBAAE,OAAO,IAAI,CAAC,GAAG,CAAC,qBAAqB,IAAI,GAAG,CAAC,CAAC;gBACpG,IAAI,YAAY,CAAC,IAAI,CAAC;oBAAE,OAAO,IAAI,CAAC,GAAG,CAAC,gCAAgC,IAAI,GAAG,CAAC,CAAC;gBACjF,IAAI,CAAC,QAAQ,IAAI,YAAY,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,CAAC;oBAAE,OAAO,IAAI,CAAC,GAAG,CAAC,6BAA6B,MAAM,mBAAmB,OAAO,UAAU,IAAI,GAAG,CAAC,CAAC;gBACtJ,IAAI,CAAC,GAAG,CAAC,WAAW,IAAI,GAAG,CAAC,CAAC;gBAC7B,IAAI,CAAC,QAAQ,EAAE,CAAC;oBACxB,6GAA6G;oBAC7G,6GAA6G;oBAC7G,QAAQ;oBACE,IAAI,CAAC,GAAG,CAAC,iDAAiD,CAAC,CAAC;oBAC5D,IAAI,IAAI,CAAC,UAAU,EAAE;wBAAE,IAAI,CAAC,OAAO,EAAE,CAAC;;wBACjC,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC;gBACjC,CAAC;gBACD,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM;oBAAE,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;gBAClD,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC,CAAC;YAC5C,CAAC;YAAC,OAAO,KAAK,EAAE,CAAC;gBACf,IAAI,CAAC,OAAO,GAAG,sCAAsC,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,CAAC;YAChH,CAAC;oBAAS,CAAC;gBACT,IAAI,CAAC,OAAO,EAAE,CAAC;gBACf,IAAI,CAAC,IAAI,EAAE,CAAC;YACd,CAAC;QACH,CAAC,CAAC,CAAC;IACL,CAAC;IAED;;;OAGG;IACK,MAAM,CAAC,IAAY,EAAE,QAAQ,GAAG,cAAc;QACpD,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QACvB,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,KAAK,GAAG,GAAG,CAAC,CAAC;QACzD,MAAM,KAAK,GAAG,KAAK,CAAC,IAAI,CAAC,CAAC;QAC1B,IAAI,KAAK,CAAC,IAAI,GAAG,QAAQ;YAAE,OAAO,KAAK,CAAC;QACxC,OAAO,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE;YAC7B,IAAI,MAAM,GAAG,CAAC,CAAC;YACf,KAAK,MAAM,IAAI,IAAI,KAAK;gBAAE,IAAI,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC;oBAAE,MAAM,EAAE,CAAC;YAC7D,OAAO,MAAM,GAAG,KAAK,CAAC,IAAI,IAAI,YAAY,CAAC;QAC7C,CAAC,CAAC,CAAC;IACL,CAAC;IAEO,IAAI;QACV,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC;QAC7B,IAAI,CAAC,OAAO,GAAG,EAAE,CAAC;QAClB,KAAK,MAAM,MAAM,IAAI,OAAO;YAAE,MAAM,EAAE,CAAC;IACzC,CAAC;CACF"}

package/dist/log.d.ts

@@ -0,0 +1,6 @@
+/** Logs what happens to each sentence. Without --verbose it logs nothing. */
+export type Log = (message: string) => void;
+export declare const silent: Log;
+/** To stderr (stdout is the MCP channel) and to the file, with the time. */
+export declare function fileLog(file: string): Log;
+//# sourceMappingURL=log.d.ts.map

package/dist/log.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"log.d.ts","sourceRoot":"","sources":["../src/log.ts"],"names":[],"mappings":"AAGA,6EAA6E;AAC7E,MAAM,MAAM,GAAG,GAAG,CAAC,OAAO,EAAE,MAAM,KAAK,IAAI,CAAC;AAE5C,eAAO,MAAM,MAAM,EAAE,GAAc,CAAC;AAEpC,4EAA4E;AAC5E,wBAAgB,OAAO,CAAC,IAAI,EAAE,MAAM,GAAG,GAAG,CAOzC"}

package/dist/log.js

@@ -0,0 +1,13 @@
+import { appendFileSync, mkdirSync } from 'node:fs';
+import { dirname } from 'node:path';
+export const silent = () => { };
+/** To stderr (stdout is the MCP channel) and to the file, with the time. */
+export function fileLog(file) {
+    mkdirSync(dirname(file), { recursive: true });
+    return (message) => {
+        const line = `${new Date().toISOString().slice(11, 23)} ${message}`;
+        console.error(line);
+        appendFileSync(file, line + '\n');
+    };
+}
+//# sourceMappingURL=log.js.map

package/dist/log.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"log.js","sourceRoot":"","sources":["../src/log.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,cAAc,EAAE,SAAS,EAAE,MAAM,SAAS,CAAC;AACpD,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,CAAC;AAKpC,MAAM,CAAC,MAAM,MAAM,GAAQ,GAAG,EAAE,GAAE,CAAC,CAAC;AAEpC,4EAA4E;AAC5E,MAAM,UAAU,OAAO,CAAC,IAAY;IAClC,SAAS,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;IAC9C,OAAO,CAAC,OAAO,EAAE,EAAE;QACjB,MAAM,IAAI,GAAG,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC,KAAK,CAAC,EAAE,EAAE,EAAE,CAAC,IAAI,OAAO,EAAE,CAAC;QACpE,OAAO,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;QACpB,cAAc,CAAC,IAAI,EAAE,IAAI,GAAG,IAAI,CAAC,CAAC;IACpC,CAAC,CAAC;AACJ,CAAC"}

package/dist/mic.d.ts

@@ -0,0 +1,19 @@
+/** The sample rate Whisper expects; the microphone is recorded at it directly. */
+export declare const SAMPLE_RATE = 16000;
+/** The open microphone: 16-bit PCM, mono, 16 kHz. */
+export interface Mic {
+    onData(listener: (chunk: Buffer) => void): void;
+    onExit(listener: (reason: string) => void): void;
+    stop(): void;
+}
+export type OpenMic = () => Mic;
+/** How ffmpeg opens the microphone on each operating system. */
+export declare function inputArgs(platform: NodeJS.Platform, device: string | undefined): string[];
+/** On Windows the microphone must be named: the first one ffmpeg lists. */
+export declare function findWindowsMic(ffmpeg: string): string | undefined;
+/** The computer microphone through ffmpeg. It is only open while listening. */
+export declare function createFfmpegMic({ ffmpeg, device }: {
+    ffmpeg: string;
+    device?: string;
+}): OpenMic;
+//# sourceMappingURL=mic.d.ts.map

package/dist/mic.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"mic.d.ts","sourceRoot":"","sources":["../src/mic.ts"],"names":[],"mappings":"AAEA,kFAAkF;AAClF,eAAO,MAAM,WAAW,QAAS,CAAC;AAElC,qDAAqD;AACrD,MAAM,WAAW,GAAG;IAClB,MAAM,CAAC,QAAQ,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,IAAI,GAAG,IAAI,CAAC;IAChD,MAAM,CAAC,QAAQ,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,IAAI,GAAG,IAAI,CAAC;IACjD,IAAI,IAAI,IAAI,CAAC;CACd;AAED,MAAM,MAAM,OAAO,GAAG,MAAM,GAAG,CAAC;AAEhC,gEAAgE;AAChE,wBAAgB,SAAS,CAAC,QAAQ,EAAE,MAAM,CAAC,QAAQ,EAAE,MAAM,EAAE,MAAM,GAAG,SAAS,GAAG,MAAM,EAAE,CAIzF;AAED,2EAA2E;AAC3E,wBAAgB,cAAc,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,CAMjE;AAED,+EAA+E;AAC/E,wBAAgB,eAAe,CAAC,EAAE,MAAM,EAAE,MAAM,EAAE,EAAE;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,MAAM,CAAC,EAAE,MAAM,CAAA;CAAE,GAAG,OAAO,CAwBhG"}