@olares/lares-core 0.15.6 → 0.15.9

package/voice/duplex.js

@@ -0,0 +1,44 @@
+/**
+ * One headset call link is half duplex. The newest recording owns it and
+ * invalidates synthesis that has not reached the speaker yet.
+ */
+export function createDuplexCoordinator(stopPlayback = async () => {}) {
+  let speech = null;
+
+  function beginSpeech(externalSignal) {
+    if (speech) {
+      speech.controller.abort();
+      void Promise.resolve(stopPlayback()).catch(() => {});
+    }
+    const controller = new AbortController();
+    const lease = {
+      controller,
+      active() {
+        return speech === lease && !controller.signal.aborted;
+      },
+      finish() {
+        if (speech === lease) speech = null;
+      },
+    };
+    speech = lease;
+    if (externalSignal) {
+      const abort = () => {
+        if (speech === lease) speech = null;
+        controller.abort();
+        void Promise.resolve(stopPlayback()).catch(() => {});
+      };
+      if (externalSignal.aborted) abort();
+      else externalSignal.addEventListener("abort", abort, { once: true });
+    }
+    return { signal: controller.signal, active: lease.active, finish: lease.finish };
+  }
+
+  async function interrupt() {
+    const current = speech;
+    speech = null;
+    current?.controller.abort();
+    await stopPlayback();
+  }
+
+  return { beginSpeech, interrupt };
+}

package/voice/endpoint.js

@@ -0,0 +1,57 @@
+/**
+ * Decide when a take is over from Silero's per-frame speech probability and the
+ * frame's loudness. A conference mic hears the whole room, so a frame only
+ * counts as the speaker when it is human speech and within `relative` of the
+ * loudest speech already heard in this take. Farther voices then do not hold
+ * the recording open.
+ */
+export const ENDPOINT = {
+  speechProb: 0.5,
+  // Phrase endpointing, the same default speech platforms use for a request
+  // (Azure end-of-silence is 500ms). Dictation waits 1.5–2s; a chat turn should not.
+  silenceMs: 500,
+  minSpeechMs: 250,
+  relative: 0.3,
+  floor: 0.004,
+  smoothFrames: 5,
+};
+
+export function frameRms(frame) {
+  if (!frame?.length) return 0;
+  let sum = 0;
+  for (let i = 0; i < frame.length; i += 1) sum += frame[i] * frame[i];
+  return Math.sqrt(sum / frame.length);
+}
+
+export function createSpeechEndpoint(options = {}) {
+  const o = { ...ENDPOINT, ...options };
+  const recent = [];
+  let reference = 0;
+  let speechMs = 0;
+  let quietMs = 0;
+  let ended = false;
+
+  return {
+    /** @returns {boolean} true once, on the frame that ends the take */
+    push(prob, rms, frameMs) {
+      if (ended) return false;
+      recent.push(rms);
+      if (recent.length > o.smoothFrames) recent.shift();
+      const level = recent.reduce((sum, value) => sum + value, 0) / recent.length;
+      const human = prob >= o.speechProb;
+      if (human && level > reference) reference = level;
+      const near = human && level >= Math.max(o.floor, reference * o.relative);
+      if (near) {
+        speechMs += frameMs;
+        quietMs = 0;
+      } else {
+        quietMs += frameMs;
+      }
+      if (speechMs >= o.minSpeechMs && quietMs >= o.silenceMs) {
+        ended = true;
+        return true;
+      }
+      return false;
+    },
+  };
+}