@ai-ecoverse/pi-audio 1.3.0 → 1.5.0

package/README.md

@@ -21,6 +21,8 @@ addAudioCounters(total, usage);
 
 pi-audio does not apply that patch. Hosts such as slicc-agent record monkeypatches in `patches/patches.json`.
 
+`@ai-ecoverse/pi-audio/openai` registers OpenAI and OpenRouter speech adapters when `registerOpenAI()` is called. Batch transcription posts multipart audio to `/audio/transcriptions`. OpenRouter (`provider` `openrouter`, or a base URL on `openrouter.ai`) posts JSON `input_audio` instead. Speech posts JSON to `/audio/speech` and reads the raw audio response. OpenAI can return pcm, mp3, wav, opus, or flac. OpenRouter returns pcm or mp3. Realtime transcription uses a WebSocket at `/realtime?intent=transcription` for `gpt-realtime-whisper` and `gpt-live-transcribe`. The catalog key is type, provider, and id, so the batch model keeps the id `gpt-4o-transcribe`.
+
 `convertAudio` and `convertAudioSource` convert between `AudioFormat`s: PCM16, float32, G.711 μ-law and A-law, mono and stereo, sample rate, and PCM16 WAV. Opus, MP3 and FLAC pass through only when the format does not change. `encodeAudioBase64` and `decodeAudioBase64` are the wire helpers for JSON protocols. `AudioChunk.data` stays binary. The same converter runs in an AudioWorklet from `@ai-ecoverse/pi-audio/worklet` (`AUDIO_CONVERT_PROCESSOR`, `registerAudioConvertWorklet`). Adapters declare `inputFormats` and `outputFormats` and do not resample on their own.
 
 ## Development

package/dist/chunk-speech.js

@@ -0,0 +1,135 @@
+import { blankUsage } from "./blank-usage.js";
+import { VoiceRequestError } from "./errors.js";
+import { sentencesOf } from "./sentences.js";
+import {
+  SpeechEventStream
+} from "./speech.js";
+import { addAudioCounters } from "./usage.js";
+function shiftSpan(span, offset) {
+  if ("entryId" in span) {
+    return {
+      entryId: span.entryId,
+      contentIndex: span.contentIndex,
+      start: span.start + offset,
+      end: span.end + offset
+    };
+  }
+  return { itemId: span.itemId, start: span.start + offset, end: span.end + offset };
+}
+function mergeUsage(usages) {
+  const total = blankUsage();
+  for (const usage of usages) {
+    total.input += usage.input;
+    total.output += usage.output;
+    total.cacheRead += usage.cacheRead;
+    total.cacheWrite += usage.cacheWrite;
+    total.totalTokens += usage.totalTokens;
+    total.cost.input += usage.cost.input;
+    total.cost.output += usage.cost.output;
+    total.cost.cacheRead += usage.cost.cacheRead;
+    total.cost.cacheWrite += usage.cost.cacheWrite;
+    total.cost.total += usage.cost.total;
+    if (usage.reasoning !== void 0) total.reasoning = (total.reasoning ?? 0) + usage.reasoning;
+    if (usage.cacheWrite1h !== void 0) {
+      total.cacheWrite1h = (total.cacheWrite1h ?? 0) + usage.cacheWrite1h;
+    }
+    addAudioCounters(total, usage);
+  }
+  return total;
+}
+function failure(error) {
+  if (error instanceof VoiceRequestError) {
+    return { type: "error", reason: error.reason, error: error.message };
+  }
+  const message = error instanceof Error ? error.message : String(error);
+  return { type: "error", reason: "error", error: message };
+}
+async function playSentence(speak, model, input, sentence, offset, options, stream, usages, state) {
+  let inner;
+  try {
+    inner = speak(model, { ...input, text: sentence }, options);
+  } catch (error) {
+    stream.push(failure(error));
+    return false;
+  }
+  let terminal = false;
+  for await (const event of inner) {
+    if (options?.signal?.aborted && event.type !== "error" && event.type !== "done") {
+      stream.push({ type: "error", reason: "aborted", error: "aborted" });
+      return false;
+    }
+    if (event.type === "start") {
+      if (!state.started) stream.push(event);
+      state.started = true;
+      continue;
+    }
+    if (event.type === "audio") {
+      stream.push({ ...event, span: shiftSpan(event.span, offset) });
+      continue;
+    }
+    if (event.type === "done") {
+      usages.push(event.usage);
+      terminal = true;
+      continue;
+    }
+    stream.push(event);
+    return false;
+  }
+  if (terminal) return true;
+  try {
+    usages.push(await inner.result());
+    return true;
+  } catch (error) {
+    stream.push(failure(error));
+    return false;
+  }
+}
+function speakInSentences(speak, model, input, options) {
+  const stream = new SpeechEventStream();
+  void pump(speak, model, input, options, stream);
+  return stream;
+}
+async function pump(speak, model, input, options, stream) {
+  if (options?.signal?.aborted) {
+    stream.push({ type: "error", reason: "aborted", error: "aborted" });
+    return;
+  }
+  const usages = [];
+  const state = { started: false };
+  let offset = 0;
+  let played = false;
+  try {
+    for await (const sentence of sentencesOf(input.text)) {
+      if (options?.signal?.aborted) {
+        stream.push({ type: "error", reason: "aborted", error: "aborted" });
+        return;
+      }
+      played = true;
+      const ok = await playSentence(
+        speak,
+        model,
+        input,
+        sentence,
+        offset,
+        options,
+        stream,
+        usages,
+        state
+      );
+      if (!ok) return;
+      offset += sentence.length;
+    }
+    if (!played) {
+      const text = typeof input.text === "string" ? input.text : "";
+      const ok = await playSentence(speak, model, input, text, 0, options, stream, usages, state);
+      if (!ok) return;
+    }
+    stream.push({ type: "done", usage: mergeUsage(usages) });
+  } catch (error) {
+    stream.push(failure(error));
+  }
+}
+export {
+  speakInSentences
+};
+//# sourceMappingURL=chunk-speech.js.map

package/dist/chunk-speech.js.map

@@ -0,0 +1,7 @@
+{
+  "version": 3,
+  "sources": ["../src/chunk-speech.ts"],
+  "sourcesContent": ["import type { ProviderRequestOptions, Usage } from '@earendil-works/pi-ai';\nimport type { TextSpan } from './audio.ts';\nimport { blankUsage } from './blank-usage.ts';\nimport { VoiceRequestError } from './errors.ts';\nimport type { SpeechModel } from './models.ts';\nimport { sentencesOf } from './sentences.ts';\nimport {\n  type SpeechApiProvider,\n  type SpeechEvent,\n  SpeechEventStream,\n  type SpeechInput,\n} from './speech.ts';\nimport { addAudioCounters } from './usage.ts';\n\nfunction shiftSpan(span: TextSpan, offset: number): TextSpan {\n  if ('entryId' in span) {\n    return {\n      entryId: span.entryId,\n      contentIndex: span.contentIndex,\n      start: span.start + offset,\n      end: span.end + offset,\n    };\n  }\n  return { itemId: span.itemId, start: span.start + offset, end: span.end + offset };\n}\n\nfunction mergeUsage(usages: Usage[]): Usage {\n  const total = blankUsage();\n  for (const usage of usages) {\n    total.input += usage.input;\n    total.output += usage.output;\n    total.cacheRead += usage.cacheRead;\n    total.cacheWrite += usage.cacheWrite;\n    total.totalTokens += usage.totalTokens;\n    total.cost.input += usage.cost.input;\n    total.cost.output += usage.cost.output;\n    total.cost.cacheRead += usage.cost.cacheRead;\n    total.cost.cacheWrite += usage.cost.cacheWrite;\n    total.cost.total += usage.cost.total;\n    if (usage.reasoning !== undefined) total.reasoning = (total.reasoning ?? 0) + usage.reasoning;\n    if (usage.cacheWrite1h !== undefined) {\n      total.cacheWrite1h = (total.cacheWrite1h ?? 0) + usage.cacheWrite1h;\n    }\n    addAudioCounters(total, usage);\n  }\n  return total;\n}\n\nfunction failure(error: unknown): SpeechEvent {\n  if (error instanceof VoiceRequestError) {\n    return { type: 'error', reason: error.reason, error: error.message };\n  }\n  const message = error instanceof Error ? error.message : String(error);\n  return { type: 'error', reason: 'error', error: message };\n}\n\nasync function playSentence(\n  speak: SpeechApiProvider['speak'],\n  model: SpeechModel<string>,\n  input: SpeechInput,\n  sentence: string,\n  offset: number,\n  options: ProviderRequestOptions | undefined,\n  stream: SpeechEventStream,\n  usages: Usage[],\n  state: { started: boolean }\n): Promise<boolean> {\n  let inner: SpeechEventStream;\n  try {\n    inner = speak(model, { ...input, text: sentence }, options);\n  } catch (error) {\n    stream.push(failure(error));\n    return false;\n  }\n  let terminal = false;\n  for await (const event of inner) {\n    if (options?.signal?.aborted && event.type !== 'error' && event.type !== 'done') {\n      stream.push({ type: 'error', reason: 'aborted', error: 'aborted' });\n      return false;\n    }\n    if (event.type === 'start') {\n      if (!state.started) stream.push(event);\n      state.started = true;\n      continue;\n    }\n    if (event.type === 'audio') {\n      stream.push({ ...event, span: shiftSpan(event.span, offset) });\n      continue;\n    }\n    if (event.type === 'done') {\n      usages.push(event.usage);\n      terminal = true;\n      continue;\n    }\n    stream.push(event);\n    return false;\n  }\n  if (terminal) return true;\n  try {\n    usages.push(await inner.result());\n    return true;\n  } catch (error) {\n    stream.push(failure(error));\n    return false;\n  }\n}\n\nexport function speakInSentences(\n  speak: SpeechApiProvider['speak'],\n  model: SpeechModel<string>,\n  input: SpeechInput,\n  options?: ProviderRequestOptions\n): SpeechEventStream {\n  const stream = new SpeechEventStream();\n  void pump(speak, model, input, options, stream);\n  return stream;\n}\n\nasync function pump(\n  speak: SpeechApiProvider['speak'],\n  model: SpeechModel<string>,\n  input: SpeechInput,\n  options: ProviderRequestOptions | undefined,\n  stream: SpeechEventStream\n) {\n  if (options?.signal?.aborted) {\n    stream.push({ type: 'error', reason: 'aborted', error: 'aborted' });\n    return;\n  }\n  const usages: Usage[] = [];\n  const state = { started: false };\n  let offset = 0;\n  let played = false;\n  try {\n    for await (const sentence of sentencesOf(input.text)) {\n      if (options?.signal?.aborted) {\n        stream.push({ type: 'error', reason: 'aborted', error: 'aborted' });\n        return;\n      }\n      played = true;\n      const ok = await playSentence(\n        speak,\n        model,\n        input,\n        sentence,\n        offset,\n        options,\n        stream,\n        usages,\n        state\n      );\n      if (!ok) return;\n      offset += sentence.length;\n    }\n    if (!played) {\n      const text = typeof input.text === 'string' ? input.text : '';\n      const ok = await playSentence(speak, model, input, text, 0, options, stream, usages, state);\n      if (!ok) return;\n    }\n    stream.push({ type: 'done', usage: mergeUsage(usages) });\n  } catch (error) {\n    stream.push(failure(error));\n  }\n}\n"],
+  "mappings": "AAEA,SAAS,kBAAkB;AAC3B,SAAS,yBAAyB;AAElC,SAAS,mBAAmB;AAC5B;AAAA,EAGE;AAAA,OAEK;AACP,SAAS,wBAAwB;AAEjC,SAAS,UAAU,MAAgB,QAA0B;AAC3D,MAAI,aAAa,MAAM;AACrB,WAAO;AAAA,MACL,SAAS,KAAK;AAAA,MACd,cAAc,KAAK;AAAA,MACnB,OAAO,KAAK,QAAQ;AAAA,MACpB,KAAK,KAAK,MAAM;AAAA,IAClB;AAAA,EACF;AACA,SAAO,EAAE,QAAQ,KAAK,QAAQ,OAAO,KAAK,QAAQ,QAAQ,KAAK,KAAK,MAAM,OAAO;AACnF;AAEA,SAAS,WAAW,QAAwB;AAC1C,QAAM,QAAQ,WAAW;AACzB,aAAW,SAAS,QAAQ;AAC1B,UAAM,SAAS,MAAM;AACrB,UAAM,UAAU,MAAM;AACtB,UAAM,aAAa,MAAM;AACzB,UAAM,cAAc,MAAM;AAC1B,UAAM,eAAe,MAAM;AAC3B,UAAM,KAAK,SAAS,MAAM,KAAK;AAC/B,UAAM,KAAK,UAAU,MAAM,KAAK;AAChC,UAAM,KAAK,aAAa,MAAM,KAAK;AACnC,UAAM,KAAK,cAAc,MAAM,KAAK;AACpC,UAAM,KAAK,SAAS,MAAM,KAAK;AAC/B,QAAI,MAAM,cAAc,OAAW,OAAM,aAAa,MAAM,aAAa,KAAK,MAAM;AACpF,QAAI,MAAM,iBAAiB,QAAW;AACpC,YAAM,gBAAgB,MAAM,gBAAgB,KAAK,MAAM;AAAA,IACzD;AACA,qBAAiB,OAAO,KAAK;AAAA,EAC/B;AACA,SAAO;AACT;AAEA,SAAS,QAAQ,OAA6B;AAC5C,MAAI,iBAAiB,mBAAmB;AACtC,WAAO,EAAE,MAAM,SAAS,QAAQ,MAAM,QAAQ,OAAO,MAAM,QAAQ;AAAA,EACrE;AACA,QAAM,UAAU,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;AACrE,SAAO,EAAE,MAAM,SAAS,QAAQ,SAAS,OAAO,QAAQ;AAC1D;AAEA,eAAe,aACb,OACA,OACA,OACA,UACA,QACA,SACA,QACA,QACA,OACkB;AAClB,MAAI;AACJ,MAAI;AACF,YAAQ,MAAM,OAAO,EAAE,GAAG,OAAO,MAAM,SAAS,GAAG,OAAO;AAAA,EAC5D,SAAS,OAAO;AACd,WAAO,KAAK,QAAQ,KAAK,CAAC;AAC1B,WAAO;AAAA,EACT;AACA,MAAI,WAAW;AACf,mBAAiB,SAAS,OAAO;AAC/B,QAAI,SAAS,QAAQ,WAAW,MAAM,SAAS,WAAW,MAAM,SAAS,QAAQ;AAC/E,aAAO,KAAK,EAAE,MAAM,SAAS,QAAQ,WAAW,OAAO,UAAU,CAAC;AAClE,aAAO;AAAA,IACT;AACA,QAAI,MAAM,SAAS,SAAS;AAC1B,UAAI,CAAC,MAAM,QAAS,QAAO,KAAK,KAAK;AACrC,YAAM,UAAU;AAChB;AAAA,IACF;AACA,QAAI,MAAM,SAAS,SAAS;AAC1B,aAAO,KAAK,EAAE,GAAG,OAAO,MAAM,UAAU,MAAM,MAAM,MAAM,EAAE,CAAC;AAC7D;AAAA,IACF;AACA,QAAI,MAAM,SAAS,QAAQ;AACzB,aAAO,KAAK,MAAM,KAAK;AACvB,iBAAW;AACX;AAAA,IACF;AACA,WAAO,KAAK,KAAK;AACjB,WAAO;AAAA,EACT;AACA,MAAI,SAAU,QAAO;AACrB,MAAI;AACF,WAAO,KAAK,MAAM,MAAM,OAAO,CAAC;AAChC,WAAO;AAAA,EACT,SAAS,OAAO;AACd,WAAO,KAAK,QAAQ,KAAK,CAAC;AAC1B,WAAO;AAAA,EACT;AACF;AAEO,SAAS,iBACd,OACA,OACA,OACA,SACmB;AACnB,QAAM,SAAS,IAAI,kBAAkB;AACrC,OAAK,KAAK,OAAO,OAAO,OAAO,SAAS,MAAM;AAC9C,SAAO;AACT;AAEA,eAAe,KACb,OACA,OACA,OACA,SACA,QACA;AACA,MAAI,SAAS,QAAQ,SAAS;AAC5B,WAAO,KAAK,EAAE,MAAM,SAAS,QAAQ,WAAW,OAAO,UAAU,CAAC;AAClE;AAAA,EACF;AACA,QAAM,SAAkB,CAAC;AACzB,QAAM,QAAQ,EAAE,SAAS,MAAM;AAC/B,MAAI,SAAS;AACb,MAAI,SAAS;AACb,MAAI;AACF,qBAAiB,YAAY,YAAY,MAAM,IAAI,GAAG;AACpD,UAAI,SAAS,QAAQ,SAAS;AAC5B,eAAO,KAAK,EAAE,MAAM,SAAS,QAAQ,WAAW,OAAO,UAAU,CAAC;AAClE;AAAA,MACF;AACA,eAAS;AACT,YAAM,KAAK,MAAM;AAAA,QACf;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF;AACA,UAAI,CAAC,GAAI;AACT,gBAAU,SAAS;AAAA,IACrB;AACA,QAAI,CAAC,QAAQ;AACX,YAAM,OAAO,OAAO,MAAM,SAAS,WAAW,MAAM,OAAO;AAC3D,YAAM,KAAK,MAAM,aAAa,OAAO,OAAO,OAAO,MAAM,GAAG,SAAS,QAAQ,QAAQ,KAAK;AAC1F,UAAI,CAAC,GAAI;AAAA,IACX;AACA,WAAO,KAAK,EAAE,MAAM,QAAQ,OAAO,WAAW,MAAM,EAAE,CAAC;AAAA,EACzD,SAAS,OAAO;AACd,WAAO,KAAK,QAAQ,KAAK,CAAC;AAAA,EAC5B;AACF;",
+  "names": []
+}

package/dist/index.js

@@ -24,6 +24,7 @@ import {
 import { SpeechEventStream } from "./speech.js";
 import { TranscriptionEventStream } from "./transcription.js";
 import { addAudioCounters, priceAudio } from "./usage.js";
+import { energyVad } from "./vad.js";
 export {
   SpeechEventStream,
   TranscriptionEventStream,
@@ -35,6 +36,7 @@ export {
   convertAudioSource,
   decodeAudioBase64,
   encodeAudioBase64,
+  energyVad,
   getVoiceModel,
   listVoiceModels,
   listen,

package/dist/index.js.map

@@ -1,7 +1,7 @@
 {
   "version": 3,
   "sources": ["../src/index.ts"],
-  "sourcesContent": ["export type {\n  AudioChunk,\n  AudioEncoding,\n  AudioFormat,\n  AudioSink,\n  AudioSource,\n  AudioTag,\n  CharAlignment,\n  TextSpan,\n} from './audio.ts';\nexport { resolveVoiceAuth } from './auth.ts';\nexport { blankUsage } from './blank-usage.ts';\nexport {\n  convertAudio,\n  convertAudioSource,\n  decodeAudioBase64,\n  encodeAudioBase64,\n} from './convert.ts';\nexport { VoiceRequestError } from './errors.ts';\nexport type {\n  Delivery,\n  DeliverySupport,\n  KnownRealtimeApi,\n  KnownSpeechApi,\n  KnownTranscriptionApi,\n  RealtimeApi,\n  RealtimeModel,\n  SpeechApi,\n  SpeechModel,\n  TranscriptionApi,\n  TranscriptionModel,\n  VoiceInfo,\n  VoiceModel,\n  VoiceModelType,\n} from './models.ts';\nexport {\n  getVoiceModel,\n  listVoiceModels,\n  registerVoiceModel,\n} from './models.ts';\nexport type {\n  RealtimeApiProvider,\n  RealtimeContext,\n  RealtimeEvent,\n  RealtimeOptions,\n  RealtimeSession,\n  RealtimeToken,\n} from './realtime.ts';\nexport {\n  connectRealtime,\n  listen,\n  registerRealtimeApiProvider,\n  registerSpeechApiProvider,\n  registerTranscriptionApiProvider,\n  speak,\n  transcribe,\n} from './registry.ts';\nexport type { SpeechApiProvider, SpeechEvent, SpeechInput } from './speech.ts';\nexport { SpeechEventStream } from './speech.ts';\nexport type {\n  AudioInput,\n  ListenOptions,\n  Transcript,\n  TranscriptionApiProvider,\n  TranscriptionEvent,\n  TranscriptSegment,\n} from './transcription.ts';\nexport { TranscriptionEventStream } from './transcription.ts';\nexport type {\n  AudioCost,\n  AudioCounters,\n  AudioUsage,\n  RealtimeAudioCost,\n  SpeechAudioCost,\n  TranscriptionAudioCost,\n} from './usage.ts';\nexport { addAudioCounters, priceAudio } from './usage.ts';\n"],
-  "mappings": "AAUA,SAAS,wBAAwB;AACjC,SAAS,kBAAkB;AAC3B;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OACK;AACP,SAAS,yBAAyB;AAiBlC;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,OACK;AASP;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OACK;AAEP,SAAS,yBAAyB;AASlC,SAAS,gCAAgC;AASzC,SAAS,kBAAkB,kBAAkB;",
+  "sourcesContent": ["export type {\n  AudioChunk,\n  AudioEncoding,\n  AudioFormat,\n  AudioSink,\n  AudioSource,\n  AudioTag,\n  CharAlignment,\n  TextSpan,\n} from './audio.ts';\nexport { resolveVoiceAuth } from './auth.ts';\nexport { blankUsage } from './blank-usage.ts';\nexport {\n  convertAudio,\n  convertAudioSource,\n  decodeAudioBase64,\n  encodeAudioBase64,\n} from './convert.ts';\nexport { VoiceRequestError } from './errors.ts';\nexport type {\n  Delivery,\n  DeliverySupport,\n  KnownRealtimeApi,\n  KnownSpeechApi,\n  KnownTranscriptionApi,\n  RealtimeApi,\n  RealtimeModel,\n  SpeechApi,\n  SpeechModel,\n  TranscriptionApi,\n  TranscriptionModel,\n  VoiceInfo,\n  VoiceModel,\n  VoiceModelType,\n} from './models.ts';\nexport {\n  getVoiceModel,\n  listVoiceModels,\n  registerVoiceModel,\n} from './models.ts';\nexport type {\n  RealtimeApiProvider,\n  RealtimeContext,\n  RealtimeEvent,\n  RealtimeOptions,\n  RealtimeSession,\n  RealtimeToken,\n} from './realtime.ts';\nexport {\n  connectRealtime,\n  listen,\n  registerRealtimeApiProvider,\n  registerSpeechApiProvider,\n  registerTranscriptionApiProvider,\n  speak,\n  transcribe,\n} from './registry.ts';\nexport type { SpeechApiProvider, SpeechEvent, SpeechInput } from './speech.ts';\nexport { SpeechEventStream } from './speech.ts';\nexport type {\n  AudioInput,\n  ListenOptions,\n  Transcript,\n  TranscriptionApiProvider,\n  TranscriptionEvent,\n  TranscriptSegment,\n} from './transcription.ts';\nexport { TranscriptionEventStream } from './transcription.ts';\nexport type {\n  AudioCost,\n  AudioCounters,\n  AudioUsage,\n  RealtimeAudioCost,\n  SpeechAudioCost,\n  TranscriptionAudioCost,\n} from './usage.ts';\nexport { addAudioCounters, priceAudio } from './usage.ts';\nexport type { VadHook, VadUpdate } from './vad.ts';\nexport { energyVad } from './vad.ts';\n"],
+  "mappings": "AAUA,SAAS,wBAAwB;AACjC,SAAS,kBAAkB;AAC3B;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OACK;AACP,SAAS,yBAAyB;AAiBlC;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,OACK;AASP;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OACK;AAEP,SAAS,yBAAyB;AASlC,SAAS,gCAAgC;AASzC,SAAS,kBAAkB,kBAAkB;AAE7C,SAAS,iBAAiB;",
   "names": []
 }

package/dist/openai/catalog.js

@@ -0,0 +1,126 @@
+import {
+  registerVoiceModel
+} from "../models.js";
+import { registerSpeechApiProvider, registerTranscriptionApiProvider } from "../registry.js";
+import { listenOpenAI } from "./listen.js";
+import { speakOpenAI } from "./speak.js";
+import { transcribeOpenAI } from "./transcribe.js";
+const tokenCost = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 };
+const openai = "https://api.openai.com/v1";
+const openrouter = "https://openrouter.ai/api/v1";
+const pcm24 = { encoding: "pcm16", sampleRate: 24e3, channels: 1 };
+const fileFormats = [
+  { encoding: "wav", sampleRate: 16e3, channels: 1 },
+  { encoding: "mp3", sampleRate: 16e3, channels: 1 },
+  { encoding: "flac", sampleRate: 16e3, channels: 1 }
+];
+const openAISpeechFormats = [
+  pcm24,
+  { encoding: "mp3", sampleRate: 24e3, channels: 1 },
+  { encoding: "wav", sampleRate: 24e3, channels: 1 },
+  { encoding: "opus", sampleRate: 24e3, channels: 1 },
+  { encoding: "flac", sampleRate: 24e3, channels: 1 }
+];
+const openRouterSpeechFormats = [
+  pcm24,
+  { encoding: "mp3", sampleRate: 24e3, channels: 1 }
+];
+const classicVoices = ["alloy", "ash", "coral", "echo", "fable", "onyx", "nova", "sage", "shimmer"];
+const miniVoices = [...classicVoices, "ballad", "verse", "marin", "cedar"];
+function voices(ids) {
+  return ids.map((id) => ({ id, name: id }));
+}
+function transcription(id, provider, baseUrl, perInputMinute, streaming, turnDetection, inputFormats) {
+  return {
+    id,
+    name: id,
+    api: streaming ? "openai-realtime-transcription" : "openai-transcriptions",
+    provider,
+    baseUrl,
+    type: "transcription",
+    streaming,
+    turnDetection,
+    inputFormats,
+    cost: tokenCost,
+    audioCost: { perInputMinute }
+  };
+}
+function speech(id, provider, baseUrl, outputFormats, voiceIds, audioCost) {
+  return {
+    id,
+    name: id,
+    api: "openai-speech",
+    provider,
+    baseUrl,
+    type: "speech",
+    textStreaming: false,
+    outputFormats,
+    voices: voices(voiceIds),
+    cloning: false,
+    cost: tokenCost,
+    audioCost
+  };
+}
+function openAICatalog() {
+  return [
+    transcription("whisper-1", "openai", openai, 6e-3, false, "none", fileFormats),
+    transcription("gpt-4o-transcribe", "openai", openai, 6e-3, false, "none", fileFormats),
+    transcription("gpt-4o-mini-transcribe", "openai", openai, 3e-3, false, "none", fileFormats),
+    transcription("gpt-realtime-whisper", "openai", openai, 0.017, true, "vad", [pcm24]),
+    transcription("gpt-live-transcribe", "openai", openai, 0.017, true, "none", [pcm24]),
+    transcription("openai/whisper-1", "openrouter", openrouter, 6e-3, false, "none", fileFormats),
+    transcription(
+      "openai/whisper-large-v3",
+      "openrouter",
+      openrouter,
+      6e-3,
+      false,
+      "none",
+      fileFormats
+    ),
+    speech("tts-1", "openai", openai, openAISpeechFormats, classicVoices, {
+      perThousandChars: 0.015
+    }),
+    speech("tts-1-hd", "openai", openai, openAISpeechFormats, classicVoices, {
+      perThousandChars: 0.03
+    }),
+    speech("gpt-4o-mini-tts", "openai", openai, openAISpeechFormats, miniVoices, {
+      perThousandChars: 15e-5,
+      perOutputMinute: 0.015
+    }),
+    speech(
+      "openai/gpt-4o-mini-tts-2025-12-15",
+      "openrouter",
+      openrouter,
+      openRouterSpeechFormats,
+      miniVoices,
+      { perThousandChars: 15e-5, perOutputMinute: 0.015 }
+    )
+  ];
+}
+function registerOpenAI() {
+  const models = openAICatalog();
+  const unregisterModels = models.map((model) => registerVoiceModel(model));
+  const unregisterBatch = registerTranscriptionApiProvider({
+    api: "openai-transcriptions",
+    transcribe: transcribeOpenAI
+  });
+  const unregisterRealtime = registerTranscriptionApiProvider({
+    api: "openai-realtime-transcription",
+    listen: listenOpenAI
+  });
+  const unregisterSpeech = registerSpeechApiProvider({ api: "openai-speech", speak: speakOpenAI });
+  let released = false;
+  return () => {
+    if (released) return;
+    released = true;
+    for (const unregister of unregisterModels) unregister();
+    unregisterBatch();
+    unregisterRealtime();
+    unregisterSpeech();
+  };
+}
+export {
+  registerOpenAI
+};
+//# sourceMappingURL=catalog.js.map

package/dist/openai/catalog.js.map

@@ -0,0 +1,7 @@
+{
+  "version": 3,
+  "sources": ["../../src/openai/catalog.ts"],
+  "sourcesContent": ["import type { AudioFormat } from '../audio.ts';\nimport {\n  registerVoiceModel,\n  type SpeechModel,\n  type TranscriptionModel,\n  type VoiceInfo,\n} from '../models.ts';\nimport { registerSpeechApiProvider, registerTranscriptionApiProvider } from '../registry.ts';\nimport { listenOpenAI } from './listen.ts';\nimport { speakOpenAI } from './speak.ts';\nimport { transcribeOpenAI } from './transcribe.ts';\n\nconst tokenCost = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 };\nconst openai = 'https://api.openai.com/v1';\nconst openrouter = 'https://openrouter.ai/api/v1';\n\nconst pcm24: AudioFormat = { encoding: 'pcm16', sampleRate: 24000, channels: 1 };\nconst fileFormats: AudioFormat[] = [\n  { encoding: 'wav', sampleRate: 16000, channels: 1 },\n  { encoding: 'mp3', sampleRate: 16000, channels: 1 },\n  { encoding: 'flac', sampleRate: 16000, channels: 1 },\n];\nconst openAISpeechFormats: AudioFormat[] = [\n  pcm24,\n  { encoding: 'mp3', sampleRate: 24000, channels: 1 },\n  { encoding: 'wav', sampleRate: 24000, channels: 1 },\n  { encoding: 'opus', sampleRate: 24000, channels: 1 },\n  { encoding: 'flac', sampleRate: 24000, channels: 1 },\n];\nconst openRouterSpeechFormats: AudioFormat[] = [\n  pcm24,\n  { encoding: 'mp3', sampleRate: 24000, channels: 1 },\n];\n\nconst classicVoices = ['alloy', 'ash', 'coral', 'echo', 'fable', 'onyx', 'nova', 'sage', 'shimmer'];\nconst miniVoices = [...classicVoices, 'ballad', 'verse', 'marin', 'cedar'];\n\nfunction voices(ids: string[]): VoiceInfo[] {\n  return ids.map((id) => ({ id, name: id }));\n}\n\nfunction transcription(\n  id: string,\n  provider: string,\n  baseUrl: string,\n  perInputMinute: number,\n  streaming: boolean,\n  turnDetection: TranscriptionModel<string>['turnDetection'],\n  inputFormats: AudioFormat[]\n): TranscriptionModel<string> {\n  return {\n    id,\n    name: id,\n    api: streaming ? 'openai-realtime-transcription' : 'openai-transcriptions',\n    provider,\n    baseUrl,\n    type: 'transcription',\n    streaming,\n    turnDetection,\n    inputFormats,\n    cost: tokenCost,\n    audioCost: { perInputMinute },\n  };\n}\n\nfunction speech(\n  id: string,\n  provider: string,\n  baseUrl: string,\n  outputFormats: AudioFormat[],\n  voiceIds: string[],\n  audioCost: SpeechModel<string>['audioCost']\n): SpeechModel<string> {\n  return {\n    id,\n    name: id,\n    api: 'openai-speech',\n    provider,\n    baseUrl,\n    type: 'speech',\n    textStreaming: false,\n    outputFormats,\n    voices: voices(voiceIds),\n    cloning: false,\n    cost: tokenCost,\n    audioCost,\n  };\n}\n\nfunction openAICatalog(): Array<TranscriptionModel<string> | SpeechModel<string>> {\n  return [\n    transcription('whisper-1', 'openai', openai, 0.006, false, 'none', fileFormats),\n    transcription('gpt-4o-transcribe', 'openai', openai, 0.006, false, 'none', fileFormats),\n    transcription('gpt-4o-mini-transcribe', 'openai', openai, 0.003, false, 'none', fileFormats),\n    transcription('gpt-realtime-whisper', 'openai', openai, 0.017, true, 'vad', [pcm24]),\n    transcription('gpt-live-transcribe', 'openai', openai, 0.017, true, 'none', [pcm24]),\n    transcription('openai/whisper-1', 'openrouter', openrouter, 0.006, false, 'none', fileFormats),\n    transcription(\n      'openai/whisper-large-v3',\n      'openrouter',\n      openrouter,\n      0.006,\n      false,\n      'none',\n      fileFormats\n    ),\n    speech('tts-1', 'openai', openai, openAISpeechFormats, classicVoices, {\n      perThousandChars: 0.015,\n    }),\n    speech('tts-1-hd', 'openai', openai, openAISpeechFormats, classicVoices, {\n      perThousandChars: 0.03,\n    }),\n    speech('gpt-4o-mini-tts', 'openai', openai, openAISpeechFormats, miniVoices, {\n      perThousandChars: 0.00015,\n      perOutputMinute: 0.015,\n    }),\n    speech(\n      'openai/gpt-4o-mini-tts-2025-12-15',\n      'openrouter',\n      openrouter,\n      openRouterSpeechFormats,\n      miniVoices,\n      { perThousandChars: 0.00015, perOutputMinute: 0.015 }\n    ),\n  ];\n}\n\nexport function registerOpenAI(): () => void {\n  const models = openAICatalog();\n  const unregisterModels = models.map((model) => registerVoiceModel(model));\n  const unregisterBatch = registerTranscriptionApiProvider({\n    api: 'openai-transcriptions',\n    transcribe: transcribeOpenAI,\n  });\n  const unregisterRealtime = registerTranscriptionApiProvider({\n    api: 'openai-realtime-transcription',\n    listen: listenOpenAI,\n  });\n  const unregisterSpeech = registerSpeechApiProvider({ api: 'openai-speech', speak: speakOpenAI });\n  let released = false;\n  return () => {\n    if (released) return;\n    released = true;\n    for (const unregister of unregisterModels) unregister();\n    unregisterBatch();\n    unregisterRealtime();\n    unregisterSpeech();\n  };\n}\n"],
+  "mappings": "AACA;AAAA,EACE;AAAA,OAIK;AACP,SAAS,2BAA2B,wCAAwC;AAC5E,SAAS,oBAAoB;AAC7B,SAAS,mBAAmB;AAC5B,SAAS,wBAAwB;AAEjC,MAAM,YAAY,EAAE,OAAO,GAAG,QAAQ,GAAG,WAAW,GAAG,YAAY,EAAE;AACrE,MAAM,SAAS;AACf,MAAM,aAAa;AAEnB,MAAM,QAAqB,EAAE,UAAU,SAAS,YAAY,MAAO,UAAU,EAAE;AAC/E,MAAM,cAA6B;AAAA,EACjC,EAAE,UAAU,OAAO,YAAY,MAAO,UAAU,EAAE;AAAA,EAClD,EAAE,UAAU,OAAO,YAAY,MAAO,UAAU,EAAE;AAAA,EAClD,EAAE,UAAU,QAAQ,YAAY,MAAO,UAAU,EAAE;AACrD;AACA,MAAM,sBAAqC;AAAA,EACzC;AAAA,EACA,EAAE,UAAU,OAAO,YAAY,MAAO,UAAU,EAAE;AAAA,EAClD,EAAE,UAAU,OAAO,YAAY,MAAO,UAAU,EAAE;AAAA,EAClD,EAAE,UAAU,QAAQ,YAAY,MAAO,UAAU,EAAE;AAAA,EACnD,EAAE,UAAU,QAAQ,YAAY,MAAO,UAAU,EAAE;AACrD;AACA,MAAM,0BAAyC;AAAA,EAC7C;AAAA,EACA,EAAE,UAAU,OAAO,YAAY,MAAO,UAAU,EAAE;AACpD;AAEA,MAAM,gBAAgB,CAAC,SAAS,OAAO,SAAS,QAAQ,SAAS,QAAQ,QAAQ,QAAQ,SAAS;AAClG,MAAM,aAAa,CAAC,GAAG,eAAe,UAAU,SAAS,SAAS,OAAO;AAEzE,SAAS,OAAO,KAA4B;AAC1C,SAAO,IAAI,IAAI,CAAC,QAAQ,EAAE,IAAI,MAAM,GAAG,EAAE;AAC3C;AAEA,SAAS,cACP,IACA,UACA,SACA,gBACA,WACA,eACA,cAC4B;AAC5B,SAAO;AAAA,IACL;AAAA,IACA,MAAM;AAAA,IACN,KAAK,YAAY,kCAAkC;AAAA,IACnD;AAAA,IACA;AAAA,IACA,MAAM;AAAA,IACN;AAAA,IACA;AAAA,IACA;AAAA,IACA,MAAM;AAAA,IACN,WAAW,EAAE,eAAe;AAAA,EAC9B;AACF;AAEA,SAAS,OACP,IACA,UACA,SACA,eACA,UACA,WACqB;AACrB,SAAO;AAAA,IACL;AAAA,IACA,MAAM;AAAA,IACN,KAAK;AAAA,IACL;AAAA,IACA;AAAA,IACA,MAAM;AAAA,IACN,eAAe;AAAA,IACf;AAAA,IACA,QAAQ,OAAO,QAAQ;AAAA,IACvB,SAAS;AAAA,IACT,MAAM;AAAA,IACN;AAAA,EACF;AACF;AAEA,SAAS,gBAAyE;AAChF,SAAO;AAAA,IACL,cAAc,aAAa,UAAU,QAAQ,MAAO,OAAO,QAAQ,WAAW;AAAA,IAC9E,cAAc,qBAAqB,UAAU,QAAQ,MAAO,OAAO,QAAQ,WAAW;AAAA,IACtF,cAAc,0BAA0B,UAAU,QAAQ,MAAO,OAAO,QAAQ,WAAW;AAAA,IAC3F,cAAc,wBAAwB,UAAU,QAAQ,OAAO,MAAM,OAAO,CAAC,KAAK,CAAC;AAAA,IACnF,cAAc,uBAAuB,UAAU,QAAQ,OAAO,MAAM,QAAQ,CAAC,KAAK,CAAC;AAAA,IACnF,cAAc,oBAAoB,cAAc,YAAY,MAAO,OAAO,QAAQ,WAAW;AAAA,IAC7F;AAAA,MACE;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,IACA,OAAO,SAAS,UAAU,QAAQ,qBAAqB,eAAe;AAAA,MACpE,kBAAkB;AAAA,IACpB,CAAC;AAAA,IACD,OAAO,YAAY,UAAU,QAAQ,qBAAqB,eAAe;AAAA,MACvE,kBAAkB;AAAA,IACpB,CAAC;AAAA,IACD,OAAO,mBAAmB,UAAU,QAAQ,qBAAqB,YAAY;AAAA,MAC3E,kBAAkB;AAAA,MAClB,iBAAiB;AAAA,IACnB,CAAC;AAAA,IACD;AAAA,MACE;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,EAAE,kBAAkB,OAAS,iBAAiB,MAAM;AAAA,IACtD;AAAA,EACF;AACF;AAEO,SAAS,iBAA6B;AAC3C,QAAM,SAAS,cAAc;AAC7B,QAAM,mBAAmB,OAAO,IAAI,CAAC,UAAU,mBAAmB,KAAK,CAAC;AACxE,QAAM,kBAAkB,iCAAiC;AAAA,IACvD,KAAK;AAAA,IACL,YAAY;AAAA,EACd,CAAC;AACD,QAAM,qBAAqB,iCAAiC;AAAA,IAC1D,KAAK;AAAA,IACL,QAAQ;AAAA,EACV,CAAC;AACD,QAAM,mBAAmB,0BAA0B,EAAE,KAAK,iBAAiB,OAAO,YAAY,CAAC;AAC/F,MAAI,WAAW;AACf,SAAO,MAAM;AACX,QAAI,SAAU;AACd,eAAW;AACX,eAAW,cAAc,iBAAkB,YAAW;AACtD,oBAAgB;AAChB,uBAAmB;AACnB,qBAAiB;AAAA,EACnB;AACF;",
+  "names": []
+}

package/dist/openai/index.js

@@ -0,0 +1,5 @@
+import { registerOpenAI } from "./catalog.js";
+export {
+  registerOpenAI
+};
+//# sourceMappingURL=index.js.map

package/dist/openai/index.js.map

@@ -0,0 +1,7 @@
+{
+  "version": 3,
+  "sources": ["../../src/openai/index.ts"],
+  "sourcesContent": ["export { registerOpenAI } from './catalog.ts';\n"],
+  "mappings": "AAAA,SAAS,sBAAsB;",
+  "names": []
+}

package/dist/openai/listen.js

@@ -0,0 +1,200 @@
+import { encodeAudioBase64 } from "../convert.js";
+import {
+  TranscriptionEventStream
+} from "../transcription.js";
+import { priceAudio } from "../usage.js";
+import {
+  asVoiceError,
+  isRecord,
+  numberField,
+  realtimeUrl,
+  requestHeaders,
+  stringField
+} from "./wire.js";
+function turnDetection(model) {
+  if (model.turnDetection === "vad") return { type: "server_vad" };
+  if (model.turnDetection === "semantic") return { type: "semantic_vad" };
+  return null;
+}
+function sessionBody(model, options, format) {
+  const transcription = { model: model.id };
+  if (options?.language) {
+    transcription.language = options.language;
+    transcription.languages = [options.language];
+  }
+  if (options?.vocabulary && options.vocabulary.length > 0)
+    transcription.keywords = options.vocabulary;
+  return {
+    type: "session.update",
+    session: {
+      type: "transcription",
+      audio: {
+        input: {
+          format: { type: "audio/pcm", rate: format.sampleRate },
+          transcription,
+          turn_detection: turnDetection(model)
+        }
+      }
+    }
+  };
+}
+function matches(chunk, format) {
+  return chunk.format.encoding === format.encoding && chunk.format.sampleRate === format.sampleRate && chunk.format.channels === format.channels;
+}
+async function messageText(data) {
+  if (typeof data === "string") return data;
+  if (data instanceof ArrayBuffer) return new TextDecoder().decode(data);
+  if (data instanceof Uint8Array) return new TextDecoder().decode(data);
+  if (typeof Blob !== "undefined" && data instanceof Blob) return data.text();
+  throw new Error("Unsupported realtime message");
+}
+function applyEvent(event, state, push) {
+  const type = stringField(event, "type");
+  if (type === "input_audio_buffer.speech_started") {
+    const at = numberField(event, "audio_start_ms");
+    push({ type: "speech_start", at: at === void 0 ? 0 : at });
+    return;
+  }
+  if (type === "conversation.item.input_audio_transcription.delta") {
+    const item = stringField(event, "item_id");
+    state.segmentId = item === void 0 ? "0" : item;
+    const delta = stringField(event, "delta");
+    state.partial += delta === void 0 ? "" : delta;
+    push({ type: "partial", segmentId: state.segmentId, text: state.partial });
+    return;
+  }
+  if (type === "conversation.item.input_audio_transcription.completed") {
+    const transcript = stringField(event, "transcript");
+    const text = transcript === void 0 ? state.partial : transcript;
+    state.turns.push(text);
+    state.partial = "";
+    push({ type: "final", segmentId: state.segmentId, text });
+    push({ type: "turn_end", text, at: 0 });
+    return;
+  }
+  if (type === "error") {
+    const detail = event.error;
+    const message = isRecord(detail) ? stringField(detail, "message") : void 0;
+    state.failed = true;
+    push({
+      type: "error",
+      reason: "error",
+      error: message === void 0 ? "Realtime transcription failed" : message
+    });
+  }
+}
+function listenOpenAI(model, audio, options) {
+  const stream = new TranscriptionEventStream();
+  void pumpListen(model, audio, options, stream);
+  return stream;
+}
+function openSocket(socket, fail) {
+  return new Promise((resolve) => {
+    let opened = false;
+    socket.addEventListener("open", () => {
+      opened = true;
+      resolve();
+    });
+    socket.addEventListener("error", () => {
+      fail("OpenAI realtime transcription socket failed");
+      if (!opened) resolve();
+    });
+    socket.addEventListener("close", () => {
+      if (!opened) {
+        fail("OpenAI realtime transcription socket closed");
+        resolve();
+      }
+    });
+  });
+}
+async function pumpListen(model, audio, options, stream) {
+  const state = {
+    turns: [],
+    partial: "",
+    segmentId: "0",
+    bytes: 0,
+    audioDone: false,
+    failed: false
+  };
+  let finished = false;
+  const push = (event) => {
+    if (finished) return;
+    if (event.type === "done" || event.type === "error") finished = true;
+    stream.push(event);
+  };
+  let socket;
+  let wake;
+  const wait = new Promise((resolve) => {
+    wake = resolve;
+  });
+  try {
+    if (options?.signal?.aborted) throw new DOMException("aborted", "AbortError");
+    const format = model.inputFormats[0];
+    if (!format) throw new Error("No input format");
+    requestHeaders(model, options);
+    const apiKey = options?.apiKey;
+    if (!apiKey) throw new Error(`No API key for provider: ${model.provider}`);
+    socket = new WebSocket(realtimeUrl(model.baseUrl), [
+      "realtime",
+      `openai-insecure-api-key.${apiKey}`
+    ]);
+    let incoming = Promise.resolve();
+    socket.addEventListener("message", (event) => {
+      incoming = incoming.then(() => onMessage(event.data, state, push, wake));
+    });
+    socket.addEventListener("close", () => wake?.());
+    await openSocket(socket, (message) => {
+      push({ type: "error", reason: "error", error: message });
+      wake?.();
+    });
+    if (finished) return;
+    socket.send(JSON.stringify(sessionBody(model, options, format)));
+    for await (const chunk of audio) {
+      if (finished) return;
+      if (options?.signal?.aborted) throw new DOMException("aborted", "AbortError");
+      if (!matches(chunk, format))
+        throw new Error(`Unsupported audio format ${chunk.format.encoding}`);
+      state.bytes += chunk.data.length;
+      socket.send(
+        JSON.stringify({ type: "input_audio_buffer.append", audio: encodeAudioBase64(chunk.data) })
+      );
+    }
+    state.audioDone = true;
+    if (model.turnDetection === "none") {
+      socket.send(JSON.stringify({ type: "input_audio_buffer.commit" }));
+    }
+    if (!finished && (model.turnDetection === "none" || state.turns.length === 0)) await wait;
+    if (finished) return;
+    const width = format.sampleRate * format.channels * 2;
+    const inputSeconds = width > 0 ? state.bytes / width : 0;
+    push({
+      type: "done",
+      transcript: {
+        text: state.turns.join(" "),
+        usage: priceAudio(model.audioCost, { inputSeconds })
+      }
+    });
+  } catch (error) {
+    const voice = asVoiceError(error);
+    push({ type: "error", reason: voice.reason, error: voice.error });
+  } finally {
+    socket?.close();
+  }
+}
+async function onMessage(data, state, push, wake) {
+  try {
+    const parsed = JSON.parse(await messageText(data));
+    if (!isRecord(parsed)) throw new Error("Realtime transcription event was not JSON");
+    applyEvent(parsed, state, push);
+    if (state.failed || state.audioDone && state.turns.length > 0) wake?.();
+  } catch (error) {
+    const voice = asVoiceError(error);
+    state.failed = true;
+    push({ type: "error", reason: voice.reason, error: voice.error });
+    wake?.();
+  }
+}
+export {
+  listenOpenAI
+};
+//# sourceMappingURL=listen.js.map

package/dist/openai/listen.js.map

@@ -0,0 +1,7 @@
+{
+  "version": 3,
+  "sources": ["../../src/openai/listen.ts"],
+  "sourcesContent": ["import type { AudioChunk, AudioFormat } from '../audio.ts';\nimport { encodeAudioBase64 } from '../convert.ts';\nimport type { TranscriptionModel } from '../models.ts';\nimport {\n  type ListenOptions,\n  type TranscriptionEvent,\n  TranscriptionEventStream,\n} from '../transcription.ts';\nimport { priceAudio } from '../usage.ts';\nimport {\n  asVoiceError,\n  isRecord,\n  numberField,\n  realtimeUrl,\n  requestHeaders,\n  stringField,\n} from './wire.ts';\n\ninterface TurnState {\n  turns: string[];\n  partial: string;\n  segmentId: string;\n  bytes: number;\n  audioDone: boolean;\n  failed: boolean;\n}\n\nfunction turnDetection(model: TranscriptionModel<string>): Record<string, unknown> | null {\n  if (model.turnDetection === 'vad') return { type: 'server_vad' };\n  if (model.turnDetection === 'semantic') return { type: 'semantic_vad' };\n  return null;\n}\n\nfunction sessionBody(\n  model: TranscriptionModel<string>,\n  options: ListenOptions | undefined,\n  format: AudioFormat\n) {\n  const transcription: Record<string, unknown> = { model: model.id };\n  if (options?.language) {\n    transcription.language = options.language;\n    transcription.languages = [options.language];\n  }\n  if (options?.vocabulary && options.vocabulary.length > 0)\n    transcription.keywords = options.vocabulary;\n  return {\n    type: 'session.update',\n    session: {\n      type: 'transcription',\n      audio: {\n        input: {\n          format: { type: 'audio/pcm', rate: format.sampleRate },\n          transcription,\n          turn_detection: turnDetection(model),\n        },\n      },\n    },\n  };\n}\n\nfunction matches(chunk: AudioChunk, format: AudioFormat): boolean {\n  return (\n    chunk.format.encoding === format.encoding &&\n    chunk.format.sampleRate === format.sampleRate &&\n    chunk.format.channels === format.channels\n  );\n}\n\nasync function messageText(data: unknown): Promise<string> {\n  if (typeof data === 'string') return data;\n  if (data instanceof ArrayBuffer) return new TextDecoder().decode(data);\n  if (data instanceof Uint8Array) return new TextDecoder().decode(data);\n  if (typeof Blob !== 'undefined' && data instanceof Blob) return data.text();\n  throw new Error('Unsupported realtime message');\n}\n\nfunction applyEvent(\n  event: Record<string, unknown>,\n  state: TurnState,\n  push: (event: TranscriptionEvent) => void\n) {\n  const type = stringField(event, 'type');\n  if (type === 'input_audio_buffer.speech_started') {\n    const at = numberField(event, 'audio_start_ms');\n    push({ type: 'speech_start', at: at === undefined ? 0 : at });\n    return;\n  }\n  if (type === 'conversation.item.input_audio_transcription.delta') {\n    const item = stringField(event, 'item_id');\n    state.segmentId = item === undefined ? '0' : item;\n    const delta = stringField(event, 'delta');\n    state.partial += delta === undefined ? '' : delta;\n    push({ type: 'partial', segmentId: state.segmentId, text: state.partial });\n    return;\n  }\n  if (type === 'conversation.item.input_audio_transcription.completed') {\n    const transcript = stringField(event, 'transcript');\n    const text = transcript === undefined ? state.partial : transcript;\n    state.turns.push(text);\n    state.partial = '';\n    push({ type: 'final', segmentId: state.segmentId, text });\n    push({ type: 'turn_end', text, at: 0 });\n    return;\n  }\n  if (type === 'error') {\n    const detail = event.error;\n    const message = isRecord(detail) ? stringField(detail, 'message') : undefined;\n    state.failed = true;\n    push({\n      type: 'error',\n      reason: 'error',\n      error: message === undefined ? 'Realtime transcription failed' : message,\n    });\n  }\n}\n\nexport function listenOpenAI(\n  model: TranscriptionModel<string>,\n  audio: AsyncIterable<AudioChunk>,\n  options?: ListenOptions\n): TranscriptionEventStream {\n  const stream = new TranscriptionEventStream();\n  void pumpListen(model, audio, options, stream);\n  return stream;\n}\n\nfunction openSocket(socket: WebSocket, fail: (message: string) => void): Promise<void> {\n  return new Promise((resolve) => {\n    let opened = false;\n    socket.addEventListener('open', () => {\n      opened = true;\n      resolve();\n    });\n    socket.addEventListener('error', () => {\n      fail('OpenAI realtime transcription socket failed');\n      if (!opened) resolve();\n    });\n    socket.addEventListener('close', () => {\n      if (!opened) {\n        fail('OpenAI realtime transcription socket closed');\n        resolve();\n      }\n    });\n  });\n}\n\nasync function pumpListen(\n  model: TranscriptionModel<string>,\n  audio: AsyncIterable<AudioChunk>,\n  options: ListenOptions | undefined,\n  stream: TranscriptionEventStream\n) {\n  const state: TurnState = {\n    turns: [],\n    partial: '',\n    segmentId: '0',\n    bytes: 0,\n    audioDone: false,\n    failed: false,\n  };\n  let finished = false;\n  const push = (event: TranscriptionEvent) => {\n    if (finished) return;\n    if (event.type === 'done' || event.type === 'error') finished = true;\n    stream.push(event);\n  };\n  let socket: WebSocket | undefined;\n  let wake: (() => void) | undefined;\n  const wait = new Promise<void>((resolve) => {\n    wake = resolve;\n  });\n  try {\n    if (options?.signal?.aborted) throw new DOMException('aborted', 'AbortError');\n    const format = model.inputFormats[0];\n    if (!format) throw new Error('No input format');\n    requestHeaders(model, options);\n    const apiKey = options?.apiKey;\n    if (!apiKey) throw new Error(`No API key for provider: ${model.provider}`);\n    socket = new WebSocket(realtimeUrl(model.baseUrl), [\n      'realtime',\n      `openai-insecure-api-key.${apiKey}`,\n    ]);\n    let incoming = Promise.resolve();\n    socket.addEventListener('message', (event) => {\n      incoming = incoming.then(() => onMessage(event.data, state, push, wake));\n    });\n    socket.addEventListener('close', () => wake?.());\n    await openSocket(socket, (message) => {\n      push({ type: 'error', reason: 'error', error: message });\n      wake?.();\n    });\n    if (finished) return;\n    socket.send(JSON.stringify(sessionBody(model, options, format)));\n    for await (const chunk of audio) {\n      if (finished) return;\n      if (options?.signal?.aborted) throw new DOMException('aborted', 'AbortError');\n      if (!matches(chunk, format))\n        throw new Error(`Unsupported audio format ${chunk.format.encoding}`);\n      state.bytes += chunk.data.length;\n      socket.send(\n        JSON.stringify({ type: 'input_audio_buffer.append', audio: encodeAudioBase64(chunk.data) })\n      );\n    }\n    state.audioDone = true;\n    if (model.turnDetection === 'none') {\n      socket.send(JSON.stringify({ type: 'input_audio_buffer.commit' }));\n    }\n    if (!finished && (model.turnDetection === 'none' || state.turns.length === 0)) await wait;\n    if (finished) return;\n    const width = format.sampleRate * format.channels * 2;\n    const inputSeconds = width > 0 ? state.bytes / width : 0;\n    push({\n      type: 'done',\n      transcript: {\n        text: state.turns.join(' '),\n        usage: priceAudio(model.audioCost, { inputSeconds }),\n      },\n    });\n  } catch (error) {\n    const voice = asVoiceError(error);\n    push({ type: 'error', reason: voice.reason, error: voice.error });\n  } finally {\n    socket?.close();\n  }\n}\n\nasync function onMessage(\n  data: unknown,\n  state: TurnState,\n  push: (event: TranscriptionEvent) => void,\n  wake: (() => void) | undefined\n) {\n  try {\n    const parsed = JSON.parse(await messageText(data)) as unknown;\n    if (!isRecord(parsed)) throw new Error('Realtime transcription event was not JSON');\n    applyEvent(parsed, state, push);\n    if (state.failed || (state.audioDone && state.turns.length > 0)) wake?.();\n  } catch (error) {\n    const voice = asVoiceError(error);\n    state.failed = true;\n    push({ type: 'error', reason: voice.reason, error: voice.error });\n    wake?.();\n  }\n}\n"],
+  "mappings": "AACA,SAAS,yBAAyB;AAElC;AAAA,EAGE;AAAA,OACK;AACP,SAAS,kBAAkB;AAC3B;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OACK;AAWP,SAAS,cAAc,OAAmE;AACxF,MAAI,MAAM,kBAAkB,MAAO,QAAO,EAAE,MAAM,aAAa;AAC/D,MAAI,MAAM,kBAAkB,WAAY,QAAO,EAAE,MAAM,eAAe;AACtE,SAAO;AACT;AAEA,SAAS,YACP,OACA,SACA,QACA;AACA,QAAM,gBAAyC,EAAE,OAAO,MAAM,GAAG;AACjE,MAAI,SAAS,UAAU;AACrB,kBAAc,WAAW,QAAQ;AACjC,kBAAc,YAAY,CAAC,QAAQ,QAAQ;AAAA,EAC7C;AACA,MAAI,SAAS,cAAc,QAAQ,WAAW,SAAS;AACrD,kBAAc,WAAW,QAAQ;AACnC,SAAO;AAAA,IACL,MAAM;AAAA,IACN,SAAS;AAAA,MACP,MAAM;AAAA,MACN,OAAO;AAAA,QACL,OAAO;AAAA,UACL,QAAQ,EAAE,MAAM,aAAa,MAAM,OAAO,WAAW;AAAA,UACrD;AAAA,UACA,gBAAgB,cAAc,KAAK;AAAA,QACrC;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACF;AAEA,SAAS,QAAQ,OAAmB,QAA8B;AAChE,SACE,MAAM,OAAO,aAAa,OAAO,YACjC,MAAM,OAAO,eAAe,OAAO,cACnC,MAAM,OAAO,aAAa,OAAO;AAErC;AAEA,eAAe,YAAY,MAAgC;AACzD,MAAI,OAAO,SAAS,SAAU,QAAO;AACrC,MAAI,gBAAgB,YAAa,QAAO,IAAI,YAAY,EAAE,OAAO,IAAI;AACrE,MAAI,gBAAgB,WAAY,QAAO,IAAI,YAAY,EAAE,OAAO,IAAI;AACpE,MAAI,OAAO,SAAS,eAAe,gBAAgB,KAAM,QAAO,KAAK,KAAK;AAC1E,QAAM,IAAI,MAAM,8BAA8B;AAChD;AAEA,SAAS,WACP,OACA,OACA,MACA;AACA,QAAM,OAAO,YAAY,OAAO,MAAM;AACtC,MAAI,SAAS,qCAAqC;AAChD,UAAM,KAAK,YAAY,OAAO,gBAAgB;AAC9C,SAAK,EAAE,MAAM,gBAAgB,IAAI,OAAO,SAAY,IAAI,GAAG,CAAC;AAC5D;AAAA,EACF;AACA,MAAI,SAAS,qDAAqD;AAChE,UAAM,OAAO,YAAY,OAAO,SAAS;AACzC,UAAM,YAAY,SAAS,SAAY,MAAM;AAC7C,UAAM,QAAQ,YAAY,OAAO,OAAO;AACxC,UAAM,WAAW,UAAU,SAAY,KAAK;AAC5C,SAAK,EAAE,MAAM,WAAW,WAAW,MAAM,WAAW,MAAM,MAAM,QAAQ,CAAC;AACzE;AAAA,EACF;AACA,MAAI,SAAS,yDAAyD;AACpE,UAAM,aAAa,YAAY,OAAO,YAAY;AAClD,UAAM,OAAO,eAAe,SAAY,MAAM,UAAU;AACxD,UAAM,MAAM,KAAK,IAAI;AACrB,UAAM,UAAU;AAChB,SAAK,EAAE,MAAM,SAAS,WAAW,MAAM,WAAW,KAAK,CAAC;AACxD,SAAK,EAAE,MAAM,YAAY,MAAM,IAAI,EAAE,CAAC;AACtC;AAAA,EACF;AACA,MAAI,SAAS,SAAS;AACpB,UAAM,SAAS,MAAM;AACrB,UAAM,UAAU,SAAS,MAAM,IAAI,YAAY,QAAQ,SAAS,IAAI;AACpE,UAAM,SAAS;AACf,SAAK;AAAA,MACH,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,OAAO,YAAY,SAAY,kCAAkC;AAAA,IACnE,CAAC;AAAA,EACH;AACF;AAEO,SAAS,aACd,OACA,OACA,SAC0B;AAC1B,QAAM,SAAS,IAAI,yBAAyB;AAC5C,OAAK,WAAW,OAAO,OAAO,SAAS,MAAM;AAC7C,SAAO;AACT;AAEA,SAAS,WAAW,QAAmB,MAAgD;AACrF,SAAO,IAAI,QAAQ,CAAC,YAAY;AAC9B,QAAI,SAAS;AACb,WAAO,iBAAiB,QAAQ,MAAM;AACpC,eAAS;AACT,cAAQ;AAAA,IACV,CAAC;AACD,WAAO,iBAAiB,SAAS,MAAM;AACrC,WAAK,6CAA6C;AAClD,UAAI,CAAC,OAAQ,SAAQ;AAAA,IACvB,CAAC;AACD,WAAO,iBAAiB,SAAS,MAAM;AACrC,UAAI,CAAC,QAAQ;AACX,aAAK,6CAA6C;AAClD,gBAAQ;AAAA,MACV;AAAA,IACF,CAAC;AAAA,EACH,CAAC;AACH;AAEA,eAAe,WACb,OACA,OACA,SACA,QACA;AACA,QAAM,QAAmB;AAAA,IACvB,OAAO,CAAC;AAAA,IACR,SAAS;AAAA,IACT,WAAW;AAAA,IACX,OAAO;AAAA,IACP,WAAW;AAAA,IACX,QAAQ;AAAA,EACV;AACA,MAAI,WAAW;AACf,QAAM,OAAO,CAAC,UAA8B;AAC1C,QAAI,SAAU;AACd,QAAI,MAAM,SAAS,UAAU,MAAM,SAAS,QAAS,YAAW;AAChE,WAAO,KAAK,KAAK;AAAA,EACnB;AACA,MAAI;AACJ,MAAI;AACJ,QAAM,OAAO,IAAI,QAAc,CAAC,YAAY;AAC1C,WAAO;AAAA,EACT,CAAC;AACD,MAAI;AACF,QAAI,SAAS,QAAQ,QAAS,OAAM,IAAI,aAAa,WAAW,YAAY;AAC5E,UAAM,SAAS,MAAM,aAAa,CAAC;AACnC,QAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,iBAAiB;AAC9C,mBAAe,OAAO,OAAO;AAC7B,UAAM,SAAS,SAAS;AACxB,QAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,4BAA4B,MAAM,QAAQ,EAAE;AACzE,aAAS,IAAI,UAAU,YAAY,MAAM,OAAO,GAAG;AAAA,MACjD;AAAA,MACA,2BAA2B,MAAM;AAAA,IACnC,CAAC;AACD,QAAI,WAAW,QAAQ,QAAQ;AAC/B,WAAO,iBAAiB,WAAW,CAAC,UAAU;AAC5C,iBAAW,SAAS,KAAK,MAAM,UAAU,MAAM,MAAM,OAAO,MAAM,IAAI,CAAC;AAAA,IACzE,CAAC;AACD,WAAO,iBAAiB,SAAS,MAAM,OAAO,CAAC;AAC/C,UAAM,WAAW,QAAQ,CAAC,YAAY;AACpC,WAAK,EAAE,MAAM,SAAS,QAAQ,SAAS,OAAO,QAAQ,CAAC;AACvD,aAAO;AAAA,IACT,CAAC;AACD,QAAI,SAAU;AACd,WAAO,KAAK,KAAK,UAAU,YAAY,OAAO,SAAS,MAAM,CAAC,CAAC;AAC/D,qBAAiB,SAAS,OAAO;AAC/B,UAAI,SAAU;AACd,UAAI,SAAS,QAAQ,QAAS,OAAM,IAAI,aAAa,WAAW,YAAY;AAC5E,UAAI,CAAC,QAAQ,OAAO,MAAM;AACxB,cAAM,IAAI,MAAM,4BAA4B,MAAM,OAAO,QAAQ,EAAE;AACrE,YAAM,SAAS,MAAM,KAAK;AAC1B,aAAO;AAAA,QACL,KAAK,UAAU,EAAE,MAAM,6BAA6B,OAAO,kBAAkB,MAAM,IAAI,EAAE,CAAC;AAAA,MAC5F;AAAA,IACF;AACA,UAAM,YAAY;AAClB,QAAI,MAAM,kBAAkB,QAAQ;AAClC,aAAO,KAAK,KAAK,UAAU,EAAE,MAAM,4BAA4B,CAAC,CAAC;AAAA,IACnE;AACA,QAAI,CAAC,aAAa,MAAM,kBAAkB,UAAU,MAAM,MAAM,WAAW,GAAI,OAAM;AACrF,QAAI,SAAU;AACd,UAAM,QAAQ,OAAO,aAAa,OAAO,WAAW;AACpD,UAAM,eAAe,QAAQ,IAAI,MAAM,QAAQ,QAAQ;AACvD,SAAK;AAAA,MACH,MAAM;AAAA,MACN,YAAY;AAAA,QACV,MAAM,MAAM,MAAM,KAAK,GAAG;AAAA,QAC1B,OAAO,WAAW,MAAM,WAAW,EAAE,aAAa,CAAC;AAAA,MACrD;AAAA,IACF,CAAC;AAAA,EACH,SAAS,OAAO;AACd,UAAM,QAAQ,aAAa,KAAK;AAChC,SAAK,EAAE,MAAM,SAAS,QAAQ,MAAM,QAAQ,OAAO,MAAM,MAAM,CAAC;AAAA,EAClE,UAAE;AACA,YAAQ,MAAM;AAAA,EAChB;AACF;AAEA,eAAe,UACb,MACA,OACA,MACA,MACA;AACA,MAAI;AACF,UAAM,SAAS,KAAK,MAAM,MAAM,YAAY,IAAI,CAAC;AACjD,QAAI,CAAC,SAAS,MAAM,EAAG,OAAM,IAAI,MAAM,2CAA2C;AAClF,eAAW,QAAQ,OAAO,IAAI;AAC9B,QAAI,MAAM,UAAW,MAAM,aAAa,MAAM,MAAM,SAAS,EAAI,QAAO;AAAA,EAC1E,SAAS,OAAO;AACd,UAAM,QAAQ,aAAa,KAAK;AAChC,UAAM,SAAS;AACf,SAAK,EAAE,MAAM,SAAS,QAAQ,MAAM,QAAQ,OAAO,MAAM,MAAM,CAAC;AAChE,WAAO;AAAA,EACT;AACF;",
+  "names": []
+}