agent-voice 0.1.3 → 0.2.1

package/dist/ask-A32EH5QX.js

@@ -0,0 +1,208 @@
+#!/usr/bin/env node
+import {
+  createRealtimeSession
+} from "./chunk-UYBFONQE.js";
+import {
+  DEFAULT_VOICE,
+  SAMPLE_RATE
+} from "./chunk-AHLLYIEW.js";
+
+// src/ask.ts
+import { createRequire } from "module";
+var require2 = createRequire(import.meta.url);
+async function ask(message, options = {}) {
+  const {
+    voice = DEFAULT_VOICE,
+    timeout = 30,
+    ack = false,
+    auth,
+    onAudioFrameSent,
+    onAssistantAudio,
+    onMicAudio
+  } = options;
+  const { AudioEngine } = require2("agent-voice-audio");
+  const streamDelayMs = Number.parseInt(
+    process.env.AGENT_VOICE_AEC_STREAM_DELAY_MS ?? "30",
+    10
+  );
+  const engine = new AudioEngine({
+    sampleRate: SAMPLE_RATE,
+    channels: 1,
+    enableAec: true,
+    streamDelayMs
+  });
+  engine.start();
+  const debug = process.env.AGENT_VOICE_DEBUG_ASK_EVENTS === "1";
+  const startMs = Date.now();
+  function logEvent(event, detail) {
+    if (!debug) return;
+    const elapsed = Date.now() - startMs;
+    const suffix = detail ? ` ${detail}` : "";
+    process.stderr.write(`[ask ${elapsed}ms] ${event}${suffix}
+`);
+  }
+  logEvent("start");
+  return new Promise((resolve, reject) => {
+    let transcript = "";
+    let timeoutTimer = null;
+    let responseStartTimer = null;
+    let transcriptTimer = null;
+    let capturePollTimer = null;
+    let speechDetected = false;
+    let initialResponseDone = false;
+    let heardAssistantAudio = false;
+    let lastAssistantAudioAt = 0;
+    let cleaned = false;
+    let settled = false;
+    async function cleanup() {
+      if (cleaned) return;
+      cleaned = true;
+      logEvent("cleanup:start");
+      if (timeoutTimer) clearTimeout(timeoutTimer);
+      if (responseStartTimer) clearTimeout(responseStartTimer);
+      if (transcriptTimer) clearTimeout(transcriptTimer);
+      if (capturePollTimer) clearInterval(capturePollTimer);
+      try {
+        engine.stop();
+        engine.close();
+      } catch {
+      }
+      session.close();
+      logEvent("cleanup:done");
+    }
+    function resolveOnce(value) {
+      if (settled) return;
+      settled = true;
+      cleanup().then(() => resolve(value));
+    }
+    function rejectOnce(error) {
+      if (settled) return;
+      settled = true;
+      cleanup().then(() => reject(error));
+    }
+    capturePollTimer = setInterval(() => {
+      if (settled) return;
+      let rawFrames = [];
+      let processedFrames = [];
+      try {
+        rawFrames = engine.readRawCapture(64);
+        processedFrames = engine.readProcessedCapture(64);
+      } catch (err) {
+        rejectOnce(
+          new Error(
+            `audio engine capture read failed: ${err instanceof Error ? err.message : String(err)}`
+          )
+        );
+        return;
+      }
+      for (const frame of rawFrames) onMicAudio?.(frame);
+      if (!heardAssistantAudio) return;
+      for (const frame of processedFrames) {
+        onAudioFrameSent?.(frame);
+        session.sendAudio(frame);
+      }
+    }, 10);
+    const session = createRealtimeSession({
+      voice,
+      mode: "default",
+      ack,
+      auth,
+      onAudioDelta(pcm16) {
+        logEvent("realtime:audio_delta", `bytes=${pcm16.length}`);
+        heardAssistantAudio = true;
+        lastAssistantAudioAt = Date.now();
+        onAssistantAudio?.(pcm16);
+        engine.play(pcm16);
+      },
+      onTranscript(text) {
+        const echoGuardMs = Number.parseInt(
+          process.env.AGENT_VOICE_ECHO_GUARD_MS ?? "1500",
+          10
+        );
+        const sinceAssistantMs = Date.now() - lastAssistantAudioAt;
+        if (heardAssistantAudio && sinceAssistantMs < echoGuardMs) {
+          logEvent(
+            "realtime:transcript_ignored_echo_guard",
+            `since_assistant_ms=${sinceAssistantMs} text="${text}"`
+          );
+          return;
+        }
+        logEvent("realtime:transcript", `text="${text}"`);
+        if (transcriptTimer) {
+          clearTimeout(transcriptTimer);
+          transcriptTimer = null;
+        }
+        transcript = text;
+        if (!ack) resolveOnce(transcript);
+      },
+      onSpeechStarted() {
+        logEvent("realtime:speech_started");
+        speechDetected = true;
+        if (timeoutTimer) {
+          clearTimeout(timeoutTimer);
+          timeoutTimer = null;
+        }
+        if (transcriptTimer) clearTimeout(transcriptTimer);
+        transcriptTimer = setTimeout(() => {
+          logEvent("timeout:no_transcript_after_speech");
+          rejectOnce(
+            new Error(
+              `No transcript received within ${timeout}s after speech started`
+            )
+          );
+        }, timeout * 1e3);
+        if (!initialResponseDone && heardAssistantAudio) {
+          try {
+            engine.play(Buffer.alloc(0));
+          } catch {
+          }
+        }
+      },
+      onInitialResponseDone() {
+        logEvent("realtime:initial_response_done");
+        initialResponseDone = true;
+        timeoutTimer = setTimeout(() => {
+          if (!speechDetected) {
+            logEvent("timeout:no_speech");
+            rejectOnce(
+              new Error(`No speech detected within ${timeout}s timeout`)
+            );
+          }
+        }, timeout * 1e3);
+      },
+      onDone() {
+        logEvent("realtime:done");
+        if (ack) resolveOnce(transcript);
+      },
+      onError(error) {
+        logEvent("realtime:error", error);
+        rejectOnce(new Error(error));
+      }
+    });
+    session.connect().then(
+      () => {
+        logEvent("realtime:connected");
+        logEvent("realtime:send_message");
+        session.sendMessage(message);
+        responseStartTimer = setTimeout(() => {
+          if (!heardAssistantAudio) {
+            logEvent("timeout:no_assistant_audio");
+            rejectOnce(
+              new Error("No assistant audio received after sending message")
+            );
+          }
+        }, 1e4);
+      },
+      (err) => {
+        logEvent(
+          "realtime:connect_error",
+          err instanceof Error ? err.message : String(err)
+        );
+        rejectOnce(err instanceof Error ? err : new Error(String(err)));
+      }
+    );
+  });
+}
+export {
+  ask
+};

package/dist/{auth-BRJKBMOE.js → auth-KET5DNSE.js}

@@ -1,8 +1,8 @@
 #!/usr/bin/env node
 import {
   writeAuthConfig
-} from "./chunk-7ERYR6ZY.js";
-import "./chunk-D3AGL5JD.js";
+} from "./chunk-RGYWLATZ.js";
+import "./chunk-AHLLYIEW.js";
 
 // src/auth.ts
 import { input, password } from "@inquirer/prompts";

package/dist/{chunk-D3AGL5JD.js → chunk-AHLLYIEW.js}

@@ -3,6 +3,7 @@
 // src/types.ts
 var SAMPLE_RATE = 24e3;
 var CHANNELS = 1;
+var BIT_DEPTH = 16;
 var VOICES = [
   "alloy",
   "ash",
@@ -21,6 +22,7 @@ var DEFAULT_VOICE = "ash";
 export {
   SAMPLE_RATE,
   CHANNELS,
+  BIT_DEPTH,
   VOICES,
   DEFAULT_VOICE
 };

package/dist/{chunk-7ERYR6ZY.js → chunk-RGYWLATZ.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 import {
   DEFAULT_VOICE
-} from "./chunk-D3AGL5JD.js";
+} from "./chunk-AHLLYIEW.js";
 
 // src/config.ts
 import { chmodSync, mkdirSync, readFileSync, writeFileSync } from "fs";

package/dist/{chunk-AQ5LP2XD.js → chunk-UYBFONQE.js}

@@ -1,71 +1,4 @@
 #!/usr/bin/env node
-import {
-  CHANNELS,
-  SAMPLE_RATE
-} from "./chunk-D3AGL5JD.js";
-
-// src/audio.ts
-import { AudioIO, SampleFormat16Bit } from "naudiodon2";
-function createAudioPlayer() {
-  const stream = AudioIO({
-    outOptions: {
-      channelCount: CHANNELS,
-      sampleFormat: SampleFormat16Bit,
-      sampleRate: SAMPLE_RATE,
-      closeOnError: true
-    }
-  });
-  let closed = false;
-  return {
-    write(pcm16) {
-      return stream.write(pcm16);
-    },
-    start() {
-      stream.start();
-    },
-    drain() {
-      if (closed) return Promise.resolve();
-      closed = true;
-      return new Promise((resolve) => {
-        stream.quit(() => resolve());
-      });
-    },
-    close() {
-      if (closed) return;
-      closed = true;
-      stream.quit();
-    }
-  };
-}
-function createAudioRecorder() {
-  const stream = AudioIO({
-    inOptions: {
-      channelCount: CHANNELS,
-      sampleFormat: SampleFormat16Bit,
-      sampleRate: SAMPLE_RATE,
-      closeOnError: true
-    }
-  });
-  let stopped = false;
-  return {
-    onData(cb) {
-      stream.on("data", cb);
-    },
-    start() {
-      stream.start();
-    },
-    stop() {
-      if (stopped) return;
-      stopped = true;
-      stream.quit();
-    },
-    close() {
-      if (stopped) return;
-      stopped = true;
-      stream.quit();
-    }
-  };
-}
 
 // src/realtime.ts
 import { OpenAIRealtimeWS } from "openai/beta/realtime/ws";
@@ -110,6 +43,9 @@ function createRealtimeSession(options) {
       const pcm16 = Buffer.from(event.delta, "base64");
       options.onAudioDelta(pcm16);
     });
+    rt.on("response.audio.done", () => {
+      options.onAudioDone?.();
+    });
     rt.on("conversation.item.input_audio_transcription.completed", (event) => {
       options.onTranscript(event.transcript);
     });
@@ -177,7 +113,5 @@ ${text}`
 }
 
 export {
-  createAudioPlayer,
-  createAudioRecorder,
   createRealtimeSession
 };

package/dist/cli.js

@@ -3,13 +3,17 @@ import {
   resolveAuth,
   resolveVoice,
   writeVoiceConfig
-} from "./chunk-7ERYR6ZY.js";
+} from "./chunk-RGYWLATZ.js";
 import {
+  BIT_DEPTH,
+  CHANNELS,
+  SAMPLE_RATE,
   VOICES
-} from "./chunk-D3AGL5JD.js";
+} from "./chunk-AHLLYIEW.js";
 
 // src/cli.ts
-import { closeSync, openSync, writeSync } from "fs";
+import { closeSync, mkdirSync, openSync, writeFileSync, writeSync } from "fs";
+import { join } from "path";
 import { Command } from "commander";
 async function withSuppressedNativeOutput() {
   const savedStdout = openSync("/dev/fd/1", "w");
@@ -18,8 +22,8 @@ async function withSuppressedNativeOutput() {
   openSync("/dev/null", "w");
   closeSync(2);
   openSync("/dev/null", "w");
-  const { ask } = await import("./ask-6HS5WYJU.js");
-  const { say } = await import("./say-PKBQ2ZDL.js");
+  const { ask } = await import("./ask-A32EH5QX.js");
+  const { say } = await import("./say-ELJAIWUM.js");
   function writeResult(text) {
     writeSync(savedStdout, `${text}
 `);
@@ -45,10 +49,42 @@ async function getMessage(flag) {
   if (stdin) return stdin;
   throw new Error("No message provided. Use -m or pipe via stdin.");
 }
+function createWavBuffer(pcm16) {
+  const header = Buffer.alloc(44);
+  const dataSize = pcm16.length;
+  const fileSize = 36 + dataSize;
+  const byteRate = SAMPLE_RATE * CHANNELS * (BIT_DEPTH / 8);
+  const blockAlign = CHANNELS * (BIT_DEPTH / 8);
+  header.write("RIFF", 0);
+  header.writeUInt32LE(fileSize, 4);
+  header.write("WAVE", 8);
+  header.write("fmt ", 12);
+  header.writeUInt32LE(16, 16);
+  header.writeUInt16LE(1, 20);
+  header.writeUInt16LE(CHANNELS, 22);
+  header.writeUInt32LE(SAMPLE_RATE, 24);
+  header.writeUInt32LE(byteRate, 28);
+  header.writeUInt16LE(blockAlign, 32);
+  header.writeUInt16LE(BIT_DEPTH, 34);
+  header.write("data", 36);
+  header.writeUInt32LE(dataSize, 40);
+  return Buffer.concat([header, pcm16]);
+}
+function writeDebugAudio(dir, assistantChunks, micChunks, modelInputChunks) {
+  mkdirSync(dir, { recursive: true });
+  const stamp = (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-");
+  const assistantFile = join(dir, `ask-${stamp}-assistant-output.wav`);
+  const micFile = join(dir, `ask-${stamp}-mic-input.wav`);
+  const modelInputFile = join(dir, `ask-${stamp}-model-input.wav`);
+  writeFileSync(assistantFile, createWavBuffer(Buffer.concat(assistantChunks)));
+  writeFileSync(micFile, createWavBuffer(Buffer.concat(micChunks)));
+  writeFileSync(modelInputFile, createWavBuffer(Buffer.concat(modelInputChunks)));
+  return { assistantFile, micFile, modelInputFile };
+}
 var program = new Command().name("agent-voice").description("AI agent voice interaction CLI");
 program.command("auth").description("Configure API key and base URL").option("--api-url <url>", "Base URL for the API").option("--api-key <key>", "API key").option("--no-verify", "Skip API key verification").action(async (opts) => {
   try {
-    const { auth } = await import("./auth-BRJKBMOE.js");
+    const { auth } = await import("./auth-KET5DNSE.js");
     await auth({
       apiUrl: opts.apiUrl,
       apiKey: opts.apiKey,
@@ -84,8 +120,11 @@ voicesCmd.command("set <voice>").description("Set the default voice").action((vo
 `);
   process.exit(0);
 });
-program.command("ask").description("Speak a message and listen for a response").option("-m, --message <text>", "Text message to speak").option("--voice <name>", "OpenAI voice", defaultVoice).option("--timeout <seconds>", "Seconds to wait for user speech", "120").option("--ack", "Speak an acknowledgment after the user responds").action(async (opts) => {
+program.command("ask").description("Speak a message and listen for a response").option("-m, --message <text>", "Text message to speak").option("--voice <name>", "OpenAI voice", defaultVoice).option("--timeout <seconds>", "Seconds to wait for user speech", "120").option("--ack", "Speak an acknowledgment after the user responds").option("--debug-audio-dir <dir>", "Write ask audio debug WAVs to this directory").action(async (opts) => {
   const { ask, writeResult, writeError } = await withSuppressedNativeOutput();
+  const assistantChunks = [];
+  const micChunks = [];
+  const modelInputChunks = [];
   try {
     const auth = resolveAuth();
     const message = await getMessage(opts.message);
@@ -93,11 +132,45 @@ program.command("ask").description("Speak a message and listen for a response").
       voice: opts.voice,
       timeout: Number.parseInt(opts.timeout, 10),
       ack: opts.ack ?? false,
-      auth
+      auth,
+      onAssistantAudio: opts.debugAudioDir ? (pcm16) => assistantChunks.push(Buffer.from(pcm16)) : void 0,
+      onMicAudio: opts.debugAudioDir ? (pcm16) => micChunks.push(Buffer.from(pcm16)) : void 0,
+      onAudioFrameSent: opts.debugAudioDir ? (pcm16) => modelInputChunks.push(Buffer.from(pcm16)) : void 0
     });
+    if (opts.debugAudioDir) {
+      const files = writeDebugAudio(
+        opts.debugAudioDir,
+        assistantChunks,
+        micChunks,
+        modelInputChunks
+      );
+      writeError(
+        `debug audio written:
+${files.assistantFile}
+${files.micFile}
+${files.modelInputFile}`
+      );
+    }
     writeResult(transcript);
     process.exit(0);
   } catch (err) {
+    if (opts.debugAudioDir) {
+      try {
+        const files = writeDebugAudio(
+          opts.debugAudioDir,
+          assistantChunks,
+          micChunks,
+          modelInputChunks
+        );
+        writeError(
+          `debug audio written:
+${files.assistantFile}
+${files.micFile}
+${files.modelInputFile}`
+        );
+      } catch {
+      }
+    }
     writeError(`${err instanceof Error ? err.message : err}`);
     process.exit(1);
   }

package/dist/index.d.ts

@@ -1,16 +1,3 @@
-type AudioPlayer = {
-    write(pcm16: Buffer): boolean;
-    start(): void;
-    drain(): Promise<void>;
-    close(): void;
-};
-type AudioRecorder = {
-    onData(cb: (pcm16: Buffer) => void): void;
-    start(): void;
-    stop(): void;
-    close(): void;
-};
-
 type AuthConfig = {
     apiKey: string;
     baseUrl?: string;
@@ -23,15 +10,18 @@ type AskOptions = {
     timeout?: number;
     ack?: boolean;
     auth?: AuthConfig;
-    createPlayer?: () => AudioPlayer;
-    createRecorder?: () => AudioRecorder;
+    createPlayer?: unknown;
+    createRecorder?: unknown;
+    onAudioFrameSent?: (pcm16: Buffer) => void;
+    onAssistantAudio?: (pcm16: Buffer) => void;
+    onMicAudio?: (pcm16: Buffer) => void;
 };
 declare function ask(message: string, options?: AskOptions): Promise<string>;
 
 type SayOptions = {
     voice?: string;
     auth?: AuthConfig;
-    createPlayer?: () => AudioPlayer;
+    createPlayer?: unknown;
 };
 declare function say(message: string, options?: SayOptions): Promise<void>;
 

package/dist/index.js

@@ -1,120 +1,5 @@
-// src/audio.ts
-import { AudioIO, SampleFormat16Bit } from "naudiodon2";
-
-// src/types.ts
-var SAMPLE_RATE = 24e3;
-var CHANNELS = 1;
-var VOICES = [
-  "alloy",
-  "ash",
-  "ballad",
-  "coral",
-  "echo",
-  "fable",
-  "nova",
-  "onyx",
-  "sage",
-  "shimmer",
-  "verse"
-];
-var DEFAULT_VOICE = "ash";
-
-// src/audio.ts
-function createAudioPlayer() {
-  const stream = AudioIO({
-    outOptions: {
-      channelCount: CHANNELS,
-      sampleFormat: SampleFormat16Bit,
-      sampleRate: SAMPLE_RATE,
-      closeOnError: true
-    }
-  });
-  let closed = false;
-  return {
-    write(pcm16) {
-      return stream.write(pcm16);
-    },
-    start() {
-      stream.start();
-    },
-    drain() {
-      if (closed) return Promise.resolve();
-      closed = true;
-      return new Promise((resolve) => {
-        stream.quit(() => resolve());
-      });
-    },
-    close() {
-      if (closed) return;
-      closed = true;
-      stream.quit();
-    }
-  };
-}
-function createAudioRecorder() {
-  const stream = AudioIO({
-    inOptions: {
-      channelCount: CHANNELS,
-      sampleFormat: SampleFormat16Bit,
-      sampleRate: SAMPLE_RATE,
-      closeOnError: true
-    }
-  });
-  let stopped = false;
-  return {
-    onData(cb) {
-      stream.on("data", cb);
-    },
-    start() {
-      stream.start();
-    },
-    stop() {
-      if (stopped) return;
-      stopped = true;
-      stream.quit();
-    },
-    close() {
-      if (stopped) return;
-      stopped = true;
-      stream.quit();
-    }
-  };
-}
-
-// src/echo-canceller.ts
-import { EchoCanceller } from "agent-voice-aec";
-var FRAME_SIZE = 480;
-var FILTER_LENGTH = 4800;
-var FRAME_BYTES = FRAME_SIZE * 2;
-function createEchoCanceller() {
-  const aec = new EchoCanceller(FRAME_SIZE, FILTER_LENGTH, SAMPLE_RATE);
-  let playbackBuffer = Buffer.alloc(0);
-  let captureBuffer = Buffer.alloc(0);
-  return {
-    playback(pcm16) {
-      playbackBuffer = Buffer.concat([playbackBuffer, pcm16]);
-      while (playbackBuffer.length >= FRAME_BYTES) {
-        aec.playback(playbackBuffer.subarray(0, FRAME_BYTES));
-        playbackBuffer = playbackBuffer.subarray(FRAME_BYTES);
-      }
-    },
-    capture(pcm16) {
-      captureBuffer = Buffer.concat([captureBuffer, pcm16]);
-      const frames = [];
-      while (captureBuffer.length >= FRAME_BYTES) {
-        const out = aec.capture(captureBuffer.subarray(0, FRAME_BYTES));
-        frames.push(out);
-        captureBuffer = captureBuffer.subarray(FRAME_BYTES);
-      }
-      return frames;
-    },
-    reset() {
-      aec.reset();
-      playbackBuffer = Buffer.alloc(0);
-      captureBuffer = Buffer.alloc(0);
-    }
-  };
-}
+// src/ask.ts
+import { createRequire } from "module";
 
 // src/realtime.ts
 import { OpenAIRealtimeWS } from "openai/beta/realtime/ws";
@@ -159,6 +44,9 @@ function createRealtimeSession(options) {
       const pcm16 = Buffer.from(event.delta, "base64");
       options.onAudioDelta(pcm16);
     });
+    rt.on("response.audio.done", () => {
+      options.onAudioDone?.();
+    });
     rt.on("conversation.item.input_audio_transcription.completed", (event) => {
       options.onTranscript(event.transcript);
     });
@@ -225,100 +113,216 @@ ${text}`
   };
 }
 
+// src/types.ts
+var SAMPLE_RATE = 24e3;
+var VOICES = [
+  "alloy",
+  "ash",
+  "ballad",
+  "coral",
+  "echo",
+  "fable",
+  "nova",
+  "onyx",
+  "sage",
+  "shimmer",
+  "verse"
+];
+var DEFAULT_VOICE = "ash";
+
 // src/ask.ts
+var require2 = createRequire(import.meta.url);
 async function ask(message, options = {}) {
   const {
     voice = DEFAULT_VOICE,
     timeout = 30,
     ack = false,
     auth,
-    createPlayer = createAudioPlayer,
-    createRecorder = createAudioRecorder
+    onAudioFrameSent,
+    onAssistantAudio,
+    onMicAudio
   } = options;
-  const player = createPlayer();
-  player.start();
-  const echoCanceller = createEchoCanceller();
+  const { AudioEngine } = require2("agent-voice-audio");
+  const streamDelayMs = Number.parseInt(
+    process.env.AGENT_VOICE_AEC_STREAM_DELAY_MS ?? "30",
+    10
+  );
+  const engine = new AudioEngine({
+    sampleRate: SAMPLE_RATE,
+    channels: 1,
+    enableAec: true,
+    streamDelayMs
+  });
+  engine.start();
+  const debug = process.env.AGENT_VOICE_DEBUG_ASK_EVENTS === "1";
+  const startMs = Date.now();
+  function logEvent(event, detail) {
+    if (!debug) return;
+    const elapsed = Date.now() - startMs;
+    const suffix = detail ? ` ${detail}` : "";
+    process.stderr.write(`[ask ${elapsed}ms] ${event}${suffix}
+`);
+  }
+  logEvent("start");
   return new Promise((resolve, reject) => {
-    let recorder = null;
-    let recorderStarted = false;
     let transcript = "";
     let timeoutTimer = null;
+    let responseStartTimer = null;
+    let transcriptTimer = null;
+    let capturePollTimer = null;
     let speechDetected = false;
     let initialResponseDone = false;
-    let interrupted = false;
+    let heardAssistantAudio = false;
+    let lastAssistantAudioAt = 0;
     let cleaned = false;
-    let resolved = false;
+    let settled = false;
     async function cleanup() {
       if (cleaned) return;
       cleaned = true;
+      logEvent("cleanup:start");
       if (timeoutTimer) clearTimeout(timeoutTimer);
-      recorder?.stop();
-      recorder?.close();
-      await player.drain();
+      if (responseStartTimer) clearTimeout(responseStartTimer);
+      if (transcriptTimer) clearTimeout(transcriptTimer);
+      if (capturePollTimer) clearInterval(capturePollTimer);
+      try {
+        engine.stop();
+        engine.close();
+      } catch {
+      }
       session.close();
+      logEvent("cleanup:done");
     }
-    function finish() {
-      if (resolved) return;
-      resolved = true;
-      cleanup().then(() => resolve(transcript));
+    function resolveOnce(value) {
+      if (settled) return;
+      settled = true;
+      cleanup().then(() => resolve(value));
     }
-    function startRecorder() {
-      if (recorderStarted) return;
-      recorderStarted = true;
-      recorder = createRecorder();
-      recorder.onData((pcm16) => {
-        const cleaned2 = echoCanceller.capture(pcm16);
-        for (const frame of cleaned2) {
-          session.sendAudio(frame);
-        }
-      });
-      recorder.start();
+    function rejectOnce(error) {
+      if (settled) return;
+      settled = true;
+      cleanup().then(() => reject(error));
     }
+    capturePollTimer = setInterval(() => {
+      if (settled) return;
+      let rawFrames = [];
+      let processedFrames = [];
+      try {
+        rawFrames = engine.readRawCapture(64);
+        processedFrames = engine.readProcessedCapture(64);
+      } catch (err) {
+        rejectOnce(
+          new Error(
+            `audio engine capture read failed: ${err instanceof Error ? err.message : String(err)}`
+          )
+        );
+        return;
+      }
+      for (const frame of rawFrames) onMicAudio?.(frame);
+      if (!heardAssistantAudio) return;
+      for (const frame of processedFrames) {
+        onAudioFrameSent?.(frame);
+        session.sendAudio(frame);
+      }
+    }, 10);
     const session = createRealtimeSession({
       voice,
       mode: "default",
       ack,
       auth,
       onAudioDelta(pcm16) {
-        echoCanceller.playback(pcm16);
-        player.write(pcm16);
-        startRecorder();
+        logEvent("realtime:audio_delta", `bytes=${pcm16.length}`);
+        heardAssistantAudio = true;
+        lastAssistantAudioAt = Date.now();
+        onAssistantAudio?.(pcm16);
+        engine.play(pcm16);
       },
       onTranscript(text) {
+        const echoGuardMs = Number.parseInt(
+          process.env.AGENT_VOICE_ECHO_GUARD_MS ?? "1500",
+          10
+        );
+        const sinceAssistantMs = Date.now() - lastAssistantAudioAt;
+        if (heardAssistantAudio && sinceAssistantMs < echoGuardMs) {
+          logEvent(
+            "realtime:transcript_ignored_echo_guard",
+            `since_assistant_ms=${sinceAssistantMs} text="${text}"`
+          );
+          return;
+        }
+        logEvent("realtime:transcript", `text="${text}"`);
+        if (transcriptTimer) {
+          clearTimeout(transcriptTimer);
+          transcriptTimer = null;
+        }
         transcript = text;
-        if (!ack) finish();
+        if (!ack) resolveOnce(transcript);
       },
       onSpeechStarted() {
+        logEvent("realtime:speech_started");
         speechDetected = true;
         if (timeoutTimer) {
           clearTimeout(timeoutTimer);
           timeoutTimer = null;
         }
-        if (!initialResponseDone) {
-          interrupted = true;
-          player.close();
+        if (transcriptTimer) clearTimeout(transcriptTimer);
+        transcriptTimer = setTimeout(() => {
+          logEvent("timeout:no_transcript_after_speech");
+          rejectOnce(
+            new Error(
+              `No transcript received within ${timeout}s after speech started`
+            )
+          );
+        }, timeout * 1e3);
+        if (!initialResponseDone && heardAssistantAudio) {
+          try {
+            engine.play(Buffer.alloc(0));
+          } catch {
+          }
         }
       },
       onInitialResponseDone() {
+        logEvent("realtime:initial_response_done");
         initialResponseDone = true;
         timeoutTimer = setTimeout(() => {
           if (!speechDetected) {
-            cleanup();
-            reject(new Error(`No speech detected within ${timeout}s timeout`));
+            logEvent("timeout:no_speech");
+            rejectOnce(
+              new Error(`No speech detected within ${timeout}s timeout`)
+            );
           }
         }, timeout * 1e3);
       },
       onDone() {
-        if (ack) finish();
+        logEvent("realtime:done");
+        if (ack) resolveOnce(transcript);
       },
-      async onError(error) {
-        await cleanup();
-        reject(new Error(error));
+      onError(error) {
+        logEvent("realtime:error", error);
+        rejectOnce(new Error(error));
       }
     });
-    session.connect().then(() => {
-      session.sendMessage(message);
-    }, reject);
+    session.connect().then(
+      () => {
+        logEvent("realtime:connected");
+        logEvent("realtime:send_message");
+        session.sendMessage(message);
+        responseStartTimer = setTimeout(() => {
+          if (!heardAssistantAudio) {
+            logEvent("timeout:no_assistant_audio");
+            rejectOnce(
+              new Error("No assistant audio received after sending message")
+            );
+          }
+        }, 1e4);
+      },
+      (err) => {
+        logEvent(
+          "realtime:connect_error",
+          err instanceof Error ? err.message : String(err)
+        );
+        rejectOnce(err instanceof Error ? err : new Error(String(err)));
+      }
+    );
   });
 }
 
@@ -353,48 +357,78 @@ function resolveVoice() {
 }
 
 // src/say.ts
+import { createRequire as createRequire2 } from "module";
+var require3 = createRequire2(import.meta.url);
 async function say(message, options = {}) {
-  const {
-    voice = DEFAULT_VOICE,
-    auth,
-    createPlayer = createAudioPlayer
-  } = options;
-  const player = createPlayer();
-  player.start();
+  const { voice = DEFAULT_VOICE, auth } = options;
+  const { AudioEngine } = require3("agent-voice-audio");
+  const engine = new AudioEngine({
+    sampleRate: SAMPLE_RATE,
+    channels: 1,
+    enableAec: false
+  });
+  engine.start();
   return new Promise((resolve, reject) => {
     let cleaned = false;
+    let settled = false;
+    let responseDoneFallbackTimer = null;
+    let completionTailTimer = null;
     function cleanup() {
       if (cleaned) return;
       cleaned = true;
+      if (responseDoneFallbackTimer) clearTimeout(responseDoneFallbackTimer);
+      if (completionTailTimer) clearTimeout(completionTailTimer);
+      try {
+        engine.stop();
+        engine.close();
+      } catch {
+      }
       session.close();
     }
+    function resolveOnce() {
+      if (settled) return;
+      settled = true;
+      cleanup();
+      resolve();
+    }
+    function rejectOnce(error) {
+      if (settled) return;
+      settled = true;
+      cleanup();
+      reject(error);
+    }
+    function scheduleTailResolve(delayMs) {
+      if (settled) return;
+      if (completionTailTimer) clearTimeout(completionTailTimer);
+      completionTailTimer = setTimeout(() => {
+        resolveOnce();
+      }, delayMs);
+    }
     const session = createRealtimeSession({
       voice,
       mode: "say",
       ack: false,
       auth,
       onAudioDelta(pcm16) {
-        player.write(pcm16);
+        engine.play(pcm16);
+      },
+      onAudioDone() {
+        scheduleTailResolve(140);
       },
       onTranscript() {
       },
       onSpeechStarted() {
       },
-      async onInitialResponseDone() {
-        try {
-          await player.drain();
-        } catch {
-          player.close();
-        }
-        cleanup();
-        resolve();
+      onInitialResponseDone() {
+        if (responseDoneFallbackTimer) clearTimeout(responseDoneFallbackTimer);
+        responseDoneFallbackTimer = setTimeout(() => {
+          scheduleTailResolve(220);
+        }, 700);
       },
       onDone() {
       },
       onError(error) {
-        player.close();
-        cleanup();
-        reject(new Error(error));
+        rejectOnce(new Error(error));
       }
     });
     session.connect().then(() => {

package/dist/say-ELJAIWUM.js

@@ -0,0 +1,92 @@
+#!/usr/bin/env node
+import {
+  createRealtimeSession
+} from "./chunk-UYBFONQE.js";
+import {
+  DEFAULT_VOICE,
+  SAMPLE_RATE
+} from "./chunk-AHLLYIEW.js";
+
+// src/say.ts
+import { createRequire } from "module";
+var require2 = createRequire(import.meta.url);
+async function say(message, options = {}) {
+  const { voice = DEFAULT_VOICE, auth } = options;
+  const { AudioEngine } = require2("agent-voice-audio");
+  const engine = new AudioEngine({
+    sampleRate: SAMPLE_RATE,
+    channels: 1,
+    enableAec: false
+  });
+  engine.start();
+  return new Promise((resolve, reject) => {
+    let cleaned = false;
+    let settled = false;
+    let responseDoneFallbackTimer = null;
+    let completionTailTimer = null;
+    function cleanup() {
+      if (cleaned) return;
+      cleaned = true;
+      if (responseDoneFallbackTimer) clearTimeout(responseDoneFallbackTimer);
+      if (completionTailTimer) clearTimeout(completionTailTimer);
+      try {
+        engine.stop();
+        engine.close();
+      } catch {
+      }
+      session.close();
+    }
+    function resolveOnce() {
+      if (settled) return;
+      settled = true;
+      cleanup();
+      resolve();
+    }
+    function rejectOnce(error) {
+      if (settled) return;
+      settled = true;
+      cleanup();
+      reject(error);
+    }
+    function scheduleTailResolve(delayMs) {
+      if (settled) return;
+      if (completionTailTimer) clearTimeout(completionTailTimer);
+      completionTailTimer = setTimeout(() => {
+        resolveOnce();
+      }, delayMs);
+    }
+    const session = createRealtimeSession({
+      voice,
+      mode: "say",
+      ack: false,
+      auth,
+      onAudioDelta(pcm16) {
+        engine.play(pcm16);
+      },
+      onAudioDone() {
+        scheduleTailResolve(140);
+      },
+      onTranscript() {
+      },
+      onSpeechStarted() {
+      },
+      onInitialResponseDone() {
+        if (responseDoneFallbackTimer) clearTimeout(responseDoneFallbackTimer);
+        responseDoneFallbackTimer = setTimeout(() => {
+          scheduleTailResolve(220);
+        }, 700);
+      },
+      onDone() {
+      },
+      onError(error) {
+        rejectOnce(new Error(error));
+      }
+    });
+    session.connect().then(() => {
+      session.sendMessage(message);
+    }, reject);
+  });
+}
+export {
+  say
+};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-voice",
-  "version": "0.1.3",
+  "version": "0.2.1",
   "description": "CLI for AI agents to interact with humans via voice",
   "type": "module",
   "main": "./dist/index.js",
@@ -18,12 +18,11 @@
     "dist"
   ],
   "dependencies": {
+    "agent-voice-audio": "^0.2.1",
     "@inquirer/prompts": "^8.2.0",
     "commander": "^13.1.0",
-    "naudiodon2": "^2.1.0",
     "openai": "^4.96.0",
-    "ws": "^8.18.0",
-    "agent-voice-aec": "0.1.1"
+    "ws": "^8.18.0"
   },
   "devDependencies": {
     "@types/node": "^22.12.0",

package/dist/ask-6HS5WYJU.js

@@ -1,145 +0,0 @@
-#!/usr/bin/env node
-import {
-  createAudioPlayer,
-  createAudioRecorder,
-  createRealtimeSession
-} from "./chunk-AQ5LP2XD.js";
-import {
-  DEFAULT_VOICE,
-  SAMPLE_RATE
-} from "./chunk-D3AGL5JD.js";
-
-// src/echo-canceller.ts
-import { EchoCanceller } from "agent-voice-aec";
-var FRAME_SIZE = 480;
-var FILTER_LENGTH = 4800;
-var FRAME_BYTES = FRAME_SIZE * 2;
-function createEchoCanceller() {
-  const aec = new EchoCanceller(FRAME_SIZE, FILTER_LENGTH, SAMPLE_RATE);
-  let playbackBuffer = Buffer.alloc(0);
-  let captureBuffer = Buffer.alloc(0);
-  return {
-    playback(pcm16) {
-      playbackBuffer = Buffer.concat([playbackBuffer, pcm16]);
-      while (playbackBuffer.length >= FRAME_BYTES) {
-        aec.playback(playbackBuffer.subarray(0, FRAME_BYTES));
-        playbackBuffer = playbackBuffer.subarray(FRAME_BYTES);
-      }
-    },
-    capture(pcm16) {
-      captureBuffer = Buffer.concat([captureBuffer, pcm16]);
-      const frames = [];
-      while (captureBuffer.length >= FRAME_BYTES) {
-        const out = aec.capture(captureBuffer.subarray(0, FRAME_BYTES));
-        frames.push(out);
-        captureBuffer = captureBuffer.subarray(FRAME_BYTES);
-      }
-      return frames;
-    },
-    reset() {
-      aec.reset();
-      playbackBuffer = Buffer.alloc(0);
-      captureBuffer = Buffer.alloc(0);
-    }
-  };
-}
-
-// src/ask.ts
-async function ask(message, options = {}) {
-  const {
-    voice = DEFAULT_VOICE,
-    timeout = 30,
-    ack = false,
-    auth,
-    createPlayer = createAudioPlayer,
-    createRecorder = createAudioRecorder
-  } = options;
-  const player = createPlayer();
-  player.start();
-  const echoCanceller = createEchoCanceller();
-  return new Promise((resolve, reject) => {
-    let recorder = null;
-    let recorderStarted = false;
-    let transcript = "";
-    let timeoutTimer = null;
-    let speechDetected = false;
-    let initialResponseDone = false;
-    let interrupted = false;
-    let cleaned = false;
-    let resolved = false;
-    async function cleanup() {
-      if (cleaned) return;
-      cleaned = true;
-      if (timeoutTimer) clearTimeout(timeoutTimer);
-      recorder?.stop();
-      recorder?.close();
-      await player.drain();
-      session.close();
-    }
-    function finish() {
-      if (resolved) return;
-      resolved = true;
-      cleanup().then(() => resolve(transcript));
-    }
-    function startRecorder() {
-      if (recorderStarted) return;
-      recorderStarted = true;
-      recorder = createRecorder();
-      recorder.onData((pcm16) => {
-        const cleaned2 = echoCanceller.capture(pcm16);
-        for (const frame of cleaned2) {
-          session.sendAudio(frame);
-        }
-      });
-      recorder.start();
-    }
-    const session = createRealtimeSession({
-      voice,
-      mode: "default",
-      ack,
-      auth,
-      onAudioDelta(pcm16) {
-        echoCanceller.playback(pcm16);
-        player.write(pcm16);
-        startRecorder();
-      },
-      onTranscript(text) {
-        transcript = text;
-        if (!ack) finish();
-      },
-      onSpeechStarted() {
-        speechDetected = true;
-        if (timeoutTimer) {
-          clearTimeout(timeoutTimer);
-          timeoutTimer = null;
-        }
-        if (!initialResponseDone) {
-          interrupted = true;
-          player.close();
-        }
-      },
-      onInitialResponseDone() {
-        initialResponseDone = true;
-        timeoutTimer = setTimeout(() => {
-          if (!speechDetected) {
-            cleanup();
-            reject(new Error(`No speech detected within ${timeout}s timeout`));
-          }
-        }, timeout * 1e3);
-      },
-      onDone() {
-        if (ack) finish();
-      },
-      async onError(error) {
-        await cleanup();
-        reject(new Error(error));
-      }
-    });
-    session.connect().then(() => {
-      session.sendMessage(message);
-    }, reject);
-  });
-}
-export {
-  ask
-};

package/dist/say-PKBQ2ZDL.js

@@ -1,62 +0,0 @@
-#!/usr/bin/env node
-import {
-  createAudioPlayer,
-  createRealtimeSession
-} from "./chunk-AQ5LP2XD.js";
-import {
-  DEFAULT_VOICE
-} from "./chunk-D3AGL5JD.js";
-
-// src/say.ts
-async function say(message, options = {}) {
-  const {
-    voice = DEFAULT_VOICE,
-    auth,
-    createPlayer = createAudioPlayer
-  } = options;
-  const player = createPlayer();
-  player.start();
-  return new Promise((resolve, reject) => {
-    let cleaned = false;
-    function cleanup() {
-      if (cleaned) return;
-      cleaned = true;
-      session.close();
-    }
-    const session = createRealtimeSession({
-      voice,
-      mode: "say",
-      ack: false,
-      auth,
-      onAudioDelta(pcm16) {
-        player.write(pcm16);
-      },
-      onTranscript() {
-      },
-      onSpeechStarted() {
-      },
-      async onInitialResponseDone() {
-        try {
-          await player.drain();
-        } catch {
-          player.close();
-        }
-        cleanup();
-        resolve();
-      },
-      onDone() {
-      },
-      onError(error) {
-        player.close();
-        cleanup();
-        reject(new Error(error));
-      }
-    });
-    session.connect().then(() => {
-      session.sendMessage(message);
-    }, reject);
-  });
-}
-export {
-  say
-};