agent-voice 0.2.3 → 0.2.4

package/dist/{ask-OIE6HL2H.js → ask-KM3JPI36.js}

@@ -127,7 +127,19 @@ async function ask(message, options = {}) {
       if (!heardAssistantAudio) return;
       for (const frame of processedFrames) {
         const rms = pcm16Rms(frame);
-        const minSpeechRms = readEnvInt("AGENT_VOICE_MIN_SPEECH_RMS", 550);
+        const configuredMinSpeechRms = readEnvInt(
+          "AGENT_VOICE_MIN_SPEECH_RMS",
+          220
+        );
+        const relaxAfterMs = readEnvInt(
+          "AGENT_VOICE_MIN_SPEECH_RMS_RELAX_AFTER_MS",
+          500
+        );
+        const relaxedMinSpeechRms = readEnvInt(
+          "AGENT_VOICE_MIN_SPEECH_RMS_RELAXED",
+          120
+        );
+        const minSpeechRms = speechDetected && speechStartedAtMs > 0 && Date.now() - speechStartedAtMs >= relaxAfterMs ? relaxedMinSpeechRms : configuredMinSpeechRms;
         if (rms >= minSpeechRms) {
           nearEndEvidenceSeen = true;
           nearEndEvidenceAtMs = Date.now();
@@ -170,19 +182,25 @@ async function ask(message, options = {}) {
         logEvent("realtime:transcript", `text="${text}"`);
         trace("realtime:transcript", { text });
         if (speechDetected) {
-          const evidenceWindowMs = readEnvInt(
-            "AGENT_VOICE_SPEECH_EVIDENCE_WINDOW_MS",
-            1200
+          const evidencePreRollMs = readEnvInt(
+            "AGENT_VOICE_SPEECH_EVIDENCE_PREROLL_MS",
+            200
+          );
+          const evidencePostRollMs = readEnvInt(
+            "AGENT_VOICE_SPEECH_EVIDENCE_POSTROLL_MS",
+            1500
           );
-          const evidenceAgeMs = nearEndEvidenceSeen ? Math.abs(nearEndEvidenceAtMs - speechStartedAtMs) : Number.POSITIVE_INFINITY;
-          if (!nearEndEvidenceSeen || evidenceAgeMs > evidenceWindowMs) {
+          const evidenceEarliestMs = speechStartedAtMs - evidencePreRollMs;
+          const evidenceLatestMs = speechStartedAtMs + evidencePostRollMs;
+          const hasTimelyNearEndEvidence = nearEndEvidenceSeen && nearEndEvidenceAtMs >= evidenceEarliestMs && nearEndEvidenceAtMs <= evidenceLatestMs;
+          if (!hasTimelyNearEndEvidence) {
             trace("realtime:transcript_ignored_no_near_end_evidence", {
               text,
               speechStartedAtMs,
               nearEndEvidenceSeen,
               nearEndEvidenceAtMs,
-              evidenceAgeMs,
-              evidenceWindowMs
+              evidenceEarliestMs,
+              evidenceLatestMs
             });
             return;
           }

package/dist/cli.js

@@ -28,8 +28,8 @@ async function withSuppressedNativeOutput() {
   openSync("/dev/null", "w");
   closeSync(2);
   openSync("/dev/null", "w");
-  const { ask } = await import("./ask-OIE6HL2H.js");
-  const { say } = await import("./say-ZVF6EX52.js");
+  const { ask } = await import("./ask-KM3JPI36.js");
+  const { say } = await import("./say-OEQQFOCC.js");
   function writeResult(text) {
     writeSync(savedStdout, `${text}
 `);
@@ -187,14 +187,50 @@ ${files.modelInputFile}`
     process.exit(1);
   }
 });
-program.command("say").description("Speak a message without listening for a response").option("-m, --message <text>", "Text message to speak").option("--voice <name>", "OpenAI voice", defaultVoice).action(async (opts) => {
+program.command("say").description("Speak a message without listening for a response").option("-m, --message <text>", "Text message to speak").option("--voice <name>", "OpenAI voice", defaultVoice).option(
+  "--debug-audio-dir <dir>",
+  "Write say audio debug WAV to this directory"
+).action(async (opts) => {
   const { say, writeError } = await withSuppressedNativeOutput();
+  const assistantChunks = [];
   try {
     const auth = resolveAuth();
     const message = await getMessage(opts.message);
-    await say(message, { voice: opts.voice, auth });
+    await say(message, {
+      voice: opts.voice,
+      auth,
+      onAssistantAudio: opts.debugAudioDir ? (pcm16) => assistantChunks.push(Buffer.from(pcm16)) : void 0
+    });
+    if (opts.debugAudioDir) {
+      mkdirSync(opts.debugAudioDir, { recursive: true });
+      const stamp = (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-");
+      const file = join(
+        opts.debugAudioDir,
+        `say-${stamp}-assistant-output.wav`
+      );
+      writeFileSync(file, createWavBuffer(Buffer.concat(assistantChunks)));
+      writeError(`debug audio written:
+${file}`);
+    }
     process.exit(0);
   } catch (err) {
+    if (opts.debugAudioDir && assistantChunks.length > 0) {
+      try {
+        mkdirSync(opts.debugAudioDir, { recursive: true });
+        const stamp = (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-");
+        const file = join(
+          opts.debugAudioDir,
+          `say-${stamp}-assistant-output.wav`
+        );
+        writeFileSync(
+          file,
+          createWavBuffer(Buffer.concat(assistantChunks))
+        );
+        writeError(`debug audio written:
+${file}`);
+      } catch {
+      }
+    }
     writeError(`${err instanceof Error ? err.message : err}`);
     process.exit(1);
   }

package/dist/index.d.ts

@@ -91,6 +91,7 @@ type SayOptions = {
         enableAec?: boolean;
         streamDelayMs?: number;
     }) => RustAudioEngine;
+    onAssistantAudio?: (pcm16: Buffer) => void;
     onTrace?: (event: {
         atMs: number;
         event: string;

package/dist/index.js

@@ -249,7 +249,19 @@ async function ask(message, options = {}) {
       if (!heardAssistantAudio) return;
       for (const frame of processedFrames) {
         const rms = pcm16Rms(frame);
-        const minSpeechRms = readEnvInt("AGENT_VOICE_MIN_SPEECH_RMS", 550);
+        const configuredMinSpeechRms = readEnvInt(
+          "AGENT_VOICE_MIN_SPEECH_RMS",
+          220
+        );
+        const relaxAfterMs = readEnvInt(
+          "AGENT_VOICE_MIN_SPEECH_RMS_RELAX_AFTER_MS",
+          500
+        );
+        const relaxedMinSpeechRms = readEnvInt(
+          "AGENT_VOICE_MIN_SPEECH_RMS_RELAXED",
+          120
+        );
+        const minSpeechRms = speechDetected && speechStartedAtMs > 0 && Date.now() - speechStartedAtMs >= relaxAfterMs ? relaxedMinSpeechRms : configuredMinSpeechRms;
         if (rms >= minSpeechRms) {
           nearEndEvidenceSeen = true;
           nearEndEvidenceAtMs = Date.now();
@@ -292,19 +304,25 @@ async function ask(message, options = {}) {
         logEvent("realtime:transcript", `text="${text}"`);
         trace("realtime:transcript", { text });
         if (speechDetected) {
-          const evidenceWindowMs = readEnvInt(
-            "AGENT_VOICE_SPEECH_EVIDENCE_WINDOW_MS",
-            1200
+          const evidencePreRollMs = readEnvInt(
+            "AGENT_VOICE_SPEECH_EVIDENCE_PREROLL_MS",
+            200
+          );
+          const evidencePostRollMs = readEnvInt(
+            "AGENT_VOICE_SPEECH_EVIDENCE_POSTROLL_MS",
+            1500
           );
-          const evidenceAgeMs = nearEndEvidenceSeen ? Math.abs(nearEndEvidenceAtMs - speechStartedAtMs) : Number.POSITIVE_INFINITY;
-          if (!nearEndEvidenceSeen || evidenceAgeMs > evidenceWindowMs) {
+          const evidenceEarliestMs = speechStartedAtMs - evidencePreRollMs;
+          const evidenceLatestMs = speechStartedAtMs + evidencePostRollMs;
+          const hasTimelyNearEndEvidence = nearEndEvidenceSeen && nearEndEvidenceAtMs >= evidenceEarliestMs && nearEndEvidenceAtMs <= evidenceLatestMs;
+          if (!hasTimelyNearEndEvidence) {
             trace("realtime:transcript_ignored_no_near_end_evidence", {
               text,
               speechStartedAtMs,
               nearEndEvidenceSeen,
               nearEndEvidenceAtMs,
-              evidenceAgeMs,
-              evidenceWindowMs
+              evidenceEarliestMs,
+              evidenceLatestMs
             });
             return;
           }
@@ -439,6 +457,7 @@ async function say(message, options = {}) {
     auth,
     createSession,
     createAudioEngine,
+    onAssistantAudio,
     onTrace
   } = options;
   const { AudioEngine } = require3("agent-voice-audio");
@@ -460,6 +479,9 @@ async function say(message, options = {}) {
     let completionTailTimer = null;
     let drainPollTimer = null;
     let drainDeadlineTimer = null;
+    let playoutDeadlineTimer = null;
+    let firstAudioAtMs = 0;
+    let totalReceivedSamples = 0;
     function cleanup() {
       if (cleaned) return;
       cleaned = true;
@@ -467,6 +489,7 @@ async function say(message, options = {}) {
       if (completionTailTimer) clearTimeout(completionTailTimer);
       if (drainPollTimer) clearInterval(drainPollTimer);
       if (drainDeadlineTimer) clearTimeout(drainDeadlineTimer);
+      if (playoutDeadlineTimer) clearTimeout(playoutDeadlineTimer);
       try {
         engine.stop();
         engine.close();
@@ -487,6 +510,28 @@ async function say(message, options = {}) {
       cleanup();
       reject(error);
     }
+    function waitForWallClockPlayout() {
+      if (settled) return;
+      if (firstAudioAtMs <= 0 || totalReceivedSamples <= 0) {
+        resolveOnce();
+        return;
+      }
+      const expectedPlayoutMs = Math.ceil(
+        totalReceivedSamples / SAMPLE_RATE * 1e3
+      );
+      const playoutTailMs = 140;
+      const dueAtMs = firstAudioAtMs + expectedPlayoutMs + playoutTailMs;
+      const waitMs = Math.max(0, dueAtMs - Date.now());
+      trace("playout:wall_clock_wait", {
+        totalReceivedSamples,
+        expectedPlayoutMs,
+        playoutTailMs,
+        waitMs
+      });
+      playoutDeadlineTimer = setTimeout(() => {
+        resolveOnce();
+      }, waitMs);
+    }
     function waitForPlaybackDrain() {
       if (settled) return;
       if (!engine.getStats) {
@@ -520,7 +565,7 @@ async function say(message, options = {}) {
         if (pending <= 0) {
           zeroStreak += 1;
           if (zeroStreak >= 3) {
-            resolveOnce();
+            waitForWallClockPlayout();
           }
           return;
         }
@@ -529,12 +574,12 @@ async function say(message, options = {}) {
           trace("drain:no_progress_timeout", {
             pendingPlaybackSamples: pending
           });
-          resolveOnce();
+          waitForWallClockPlayout();
         }
       }, 20);
       drainDeadlineTimer = setTimeout(() => {
         trace("drain:deadline");
-        resolveOnce();
+        waitForWallClockPlayout();
       }, absoluteDeadlineMs);
     }
     function scheduleTailResolve(delayMs) {
@@ -551,7 +596,12 @@ async function say(message, options = {}) {
       ack: false,
       auth,
       onAudioDelta(pcm16) {
+        if (firstAudioAtMs <= 0) {
+          firstAudioAtMs = Date.now();
+        }
+        totalReceivedSamples += Math.floor(pcm16.length / 2);
         engine.play(pcm16);
+        onAssistantAudio?.(pcm16);
         trace("realtime:audio_delta", { bytes: pcm16.length });
       },
       onAudioDone() {

package/dist/{say-ZVF6EX52.js → say-OEQQFOCC.js}

@@ -16,6 +16,7 @@ async function say(message, options = {}) {
     auth,
     createSession,
     createAudioEngine,
+    onAssistantAudio,
     onTrace
   } = options;
   const { AudioEngine } = require2("agent-voice-audio");
@@ -37,6 +38,9 @@ async function say(message, options = {}) {
     let completionTailTimer = null;
     let drainPollTimer = null;
     let drainDeadlineTimer = null;
+    let playoutDeadlineTimer = null;
+    let firstAudioAtMs = 0;
+    let totalReceivedSamples = 0;
     function cleanup() {
       if (cleaned) return;
       cleaned = true;
@@ -44,6 +48,7 @@ async function say(message, options = {}) {
       if (completionTailTimer) clearTimeout(completionTailTimer);
       if (drainPollTimer) clearInterval(drainPollTimer);
       if (drainDeadlineTimer) clearTimeout(drainDeadlineTimer);
+      if (playoutDeadlineTimer) clearTimeout(playoutDeadlineTimer);
       try {
         engine.stop();
         engine.close();
@@ -64,6 +69,28 @@ async function say(message, options = {}) {
       cleanup();
       reject(error);
     }
+    function waitForWallClockPlayout() {
+      if (settled) return;
+      if (firstAudioAtMs <= 0 || totalReceivedSamples <= 0) {
+        resolveOnce();
+        return;
+      }
+      const expectedPlayoutMs = Math.ceil(
+        totalReceivedSamples / SAMPLE_RATE * 1e3
+      );
+      const playoutTailMs = 140;
+      const dueAtMs = firstAudioAtMs + expectedPlayoutMs + playoutTailMs;
+      const waitMs = Math.max(0, dueAtMs - Date.now());
+      trace("playout:wall_clock_wait", {
+        totalReceivedSamples,
+        expectedPlayoutMs,
+        playoutTailMs,
+        waitMs
+      });
+      playoutDeadlineTimer = setTimeout(() => {
+        resolveOnce();
+      }, waitMs);
+    }
     function waitForPlaybackDrain() {
       if (settled) return;
       if (!engine.getStats) {
@@ -97,7 +124,7 @@ async function say(message, options = {}) {
         if (pending <= 0) {
           zeroStreak += 1;
           if (zeroStreak >= 3) {
-            resolveOnce();
+            waitForWallClockPlayout();
           }
           return;
         }
@@ -106,12 +133,12 @@ async function say(message, options = {}) {
           trace("drain:no_progress_timeout", {
             pendingPlaybackSamples: pending
           });
-          resolveOnce();
+          waitForWallClockPlayout();
         }
       }, 20);
       drainDeadlineTimer = setTimeout(() => {
         trace("drain:deadline");
-        resolveOnce();
+        waitForWallClockPlayout();
       }, absoluteDeadlineMs);
     }
     function scheduleTailResolve(delayMs) {
@@ -128,7 +155,12 @@ async function say(message, options = {}) {
       ack: false,
       auth,
       onAudioDelta(pcm16) {
+        if (firstAudioAtMs <= 0) {
+          firstAudioAtMs = Date.now();
+        }
+        totalReceivedSamples += Math.floor(pcm16.length / 2);
         engine.play(pcm16);
+        onAssistantAudio?.(pcm16);
         trace("realtime:audio_delta", { bytes: pcm16.length });
       },
       onAudioDone() {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-voice",
-  "version": "0.2.3",
+  "version": "0.2.4",
   "description": "CLI for AI agents to interact with humans via voice",
   "type": "module",
   "main": "./dist/index.js",