@jskit-ai/assistant-voice-sherpa 0.1.13 → 0.1.15

package/README.md

@@ -159,10 +159,18 @@ model paths. Speaker IDs are specific to the selected model. Unknown IDs and
 invalid/duplicate catalogue entries fail explicitly. Voices do not change a
 conversation's reasoning model. Tail padding supplies recognizer lookahead when a
 recording ends immediately after a word; it does not wait for wall-clock silence.
-The RMS floor prevents silent input from starting a recognition utterance. After
-an audible onset, all frames reach the recognizer until the utterance ends. Its
-default is 0.001 (−60 dBFS); configure a positive value up to 0.1 for the input
-device. This is an input-level check, not a classifier for speech versus noise.
+Its default is 0.45 seconds; an explicit value from 0 to 3 seconds overrides it.
+The RMS floor is disabled by default (`0`), preserving native silence handling
+and finalization. An explicit positive value up to 0.1 gates the initial input;
+`0.001` (−60 dBFS) is the opt-in example above. After onset, all frames reach the
+recognizer until the utterance ends. With a floor enabled, finalizing an utterance
+that never reached the floor returns no words without feeding native lookahead.
+This is an input-level check, not a classifier for speech versus noise.
+
+The synthesis worker has a separate 60-second operation watchdog for model load
+and generation. A timeout fails the operation and releases the worker process;
+it is a worker lifecycle safeguard, not the microphone's recording limit or an
+extension of the original recognition timing policy.
 
 For separate models, use `synthesizers: { modelId: nativeConfiguration }` instead
 of `synthesizer`, and give each voice a matching `modelId`. Speaker IDs are checked

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@jskit-ai/assistant-voice-sherpa",
-  "version": "0.1.13",
+  "version": "0.1.15",
   "type": "module",
   "description": "Optional native Sherpa speech engine and explicit model preparation for assistant voice.",
   "scripts": {
@@ -19,7 +19,7 @@
     "tooling"
   ],
   "dependencies": {
-    "@jskit-ai/assistant-voice": "0.1.13",
+    "@jskit-ai/assistant-voice": "0.1.15",
     "sherpa-onnx-node": "1.13.6"
   },
   "jskit": {

package/src/sherpaSpeechEngine.js

@@ -1,6 +1,6 @@
 import path from "node:path";
 import { createRequire } from "node:module";
-import { createSynthesisProcess } from "./synthesisProcess.js";
+import { createSynthesisProcess, SYNTHESIS_REQUEST_TIMEOUT_MS } from "./synthesisProcess.js";
 
 import {
   VOICE_INPUT_SAMPLE_RATE,
@@ -53,13 +53,13 @@ async function createSherpaSpeechEngine({
     return value;
   }
   configuration = modelPaths(configuration);
-  const tailPaddingSeconds = configuration.recognizerTailPaddingSeconds ?? 1;
+  const tailPaddingSeconds = configuration.recognizerTailPaddingSeconds ?? 0.45;
   if (!Number.isFinite(tailPaddingSeconds) || tailPaddingSeconds < 0 || tailPaddingSeconds > 3) {
     throw new TypeError("Recognizer tail padding must be between zero and three seconds.");
   }
-  const minimumRms = configuration.recognizerMinimumRms ?? 0.001;
-  if (!Number.isFinite(minimumRms) || minimumRms <= 0 || minimumRms > 0.1) {
-    throw new TypeError("Recognizer minimum RMS must be greater than zero and at most 0.1.");
+  const minimumRms = configuration.recognizerMinimumRms ?? 0;
+  if (!Number.isFinite(minimumRms) || minimumRms < 0 || minimumRms > 0.1) {
+    throw new TypeError("Recognizer minimum RMS must be between zero and 0.1.");
   }
   const recognizer = new runtime.OnlineRecognizer(configuration.recognizer || {
     decodingMethod: "modified_beam_search",
@@ -177,9 +177,9 @@ async function createSherpaSpeechEngine({
         return latestText;
       }
       const samples = pcm16LeToFloat32(frame);
-      // This recognizer can invent words from silent input. Wait for an audible
-      // onset; once it starts, keep every frame for pauses and native endpoints.
-      if (!acceptedSamples) {
+      // An explicit RMS floor gates onset only; after onset, retain every frame
+      // for pauses and native endpoints. The default retains all input.
+      if (minimumRms > 0 && !acceptedSamples) {
         let energy = 0;
         for (const sample of samples) energy += sample * sample;
         if (!samples.length || Math.sqrt(energy / samples.length) < minimumRms) return "";
@@ -216,7 +216,7 @@ async function createSherpaSpeechEngine({
         return latestText;
       }
       closed = true;
-      if (!acceptedSamples) return "";
+      if (minimumRms > 0 && !acceptedSamples) return "";
       stream.acceptWaveform({
         sampleRate: VOICE_INPUT_SAMPLE_RATE,
         // Supply the model's lookahead when push-to-talk ends on the final word.
@@ -288,6 +288,7 @@ async function createSherpaSpeechEngine({
       version: runtimeValue(runtime, "version")
     }),
     sampleRate,
+    synthesisTimeoutMs: 2 * SYNTHESIS_REQUEST_TIMEOUT_MS,
     synthesize,
     warmup
   });

package/src/synthesisProcess.js

@@ -1,5 +1,7 @@
 import { fork } from "node:child_process";
 
+export const SYNTHESIS_REQUEST_TIMEOUT_MS = 60_000;
+
 // Sherpa's Node API has no explicit model disposal. Process exit releases both
 // the model and ONNX's native arenas before another voice is loaded.
 export async function createSynthesisProcess(configuration, {
@@ -46,7 +48,7 @@ export async function createSynthesisProcess(configuration, {
     if (stopped) return Promise.reject(new Error("Speech worker is unavailable."));
     if (pending) return Promise.reject(new Error("Speech synthesis is already running."));
     return new Promise((resolve, reject) => {
-      const timer = setTimeout(() => { settle(new Error("Speech worker timed out.")); void close(); }, 60_000);
+      const timer = setTimeout(() => { settle(new Error("Speech worker timed out.")); void close(); }, SYNTHESIS_REQUEST_TIMEOUT_MS);
       pending = { resolve, reject, onAudio, timer };
       child.send({ type, ...fields }, error => { if (error) settle(error); });
     });