getpatter 0.6.1 → 0.6.2

package/dist/index.mjs

@@ -6,6 +6,7 @@ import {
   CallMetricsAccumulator,
   DEFAULT_MIN_SENTENCE_LEN,
   DEFAULT_PRICING,
+  DeepgramModel,
   DeepgramSTT,
   DefaultToolExecutor,
   ElevenLabsConvAIAdapter,
@@ -15,12 +16,12 @@ import {
   LLMLoop,
   MetricsStore,
   OpenAILLMProvider,
-  OpenAIRealtime2Adapter,
-  OpenAIRealtimeAdapter,
+  PRICING_LAST_UPDATED,
+  PRICING_VERSION,
   PatterConnectionError,
   PatterError,
-  PcmCarry,
   PipelineHookExecutor,
+  PricingUnit,
   ProvisionError,
   RateLimitError,
   RemoteMessageHandler,
@@ -32,18 +33,14 @@ import {
   SPAN_TOOL,
   SPAN_TTS,
   SentenceChunker,
-  StatefulResampler,
   TestSession,
+  VERSION,
   calculateRealtimeCost,
   calculateSttCost,
   calculateTelephonyCost,
   calculateTtsCost,
   callsToCsv,
   callsToJson,
-  createResampler16kTo8k,
-  createResampler24kTo16k,
-  createResampler24kTo8k,
-  createResampler8kTo16k,
   initTracing,
   isRemoteUrl,
   isTracingEnabled,
@@ -53,14 +50,29 @@ import {
   mergePricing,
   mountApi,
   mountDashboard,
+  resolveLogRoot,
+  startSpan
+} from "./chunk-LE63CSOB.mjs";
+import {
+  OpenAIRealtime2Adapter,
+  OpenAIRealtimeAdapter,
+  OpenAIRealtimeAudioFormat,
+  OpenAIRealtimeModel,
+  OpenAIRealtimeVADType,
+  OpenAITranscriptionModel,
+  OpenAIVoice,
+  PcmCarry,
+  StatefulResampler,
+  createResampler16kTo8k,
+  createResampler24kTo16k,
+  createResampler24kTo8k,
+  createResampler8kTo16k,
   mulawToPcm16,
   pcm16ToMulaw,
   resample16kTo8k,
   resample24kTo16k,
-  resample8kTo16k,
-  resolveLogRoot,
-  startSpan
-} from "./chunk-TEW3NAZJ.mjs";
+  resample8kTo16k
+} from "./chunk-CL2U3YET.mjs";
 import {
   MinWordsStrategy,
   evaluateStrategies,
@@ -75,7 +87,7 @@ import {
 } from "./chunk-6GR5MHHQ.mjs";
 import {
   SileroVAD
-} from "./chunk-RV7APPYE.mjs";
+} from "./chunk-R2T4JABZ.mjs";
 import {
   __dirname,
   __require,
@@ -105,7 +117,7 @@ var Realtime = class {
       );
     }
     this.apiKey = key;
-    this.model = opts.model ?? "gpt-4o-mini-realtime-preview";
+    this.model = opts.model ?? "gpt-realtime-mini";
     this.voice = opts.voice ?? "alloy";
     this.reasoningEffort = opts.reasoningEffort;
     this.inputAudioTranscriptionModel = opts.inputAudioTranscriptionModel;
@@ -557,7 +569,9 @@ function resolvePersistRoot(persist) {
   if (persist === false) return null;
   if (persist === true) return resolveLogRoot("auto");
   if (typeof persist === "string") return resolveLogRoot(persist);
-  return resolveLogRoot();
+  const envRoot = resolveLogRoot();
+  if (envRoot !== null) return envRoot;
+  return resolveLogRoot("auto");
 }
 function closeParkedConnections(slot) {
   if (slot.stt) {
@@ -573,6 +587,11 @@ function closeParkedConnections(slot) {
     }
   }
   if (slot.openaiRealtime) {
+    const wsAny = slot.openaiRealtime;
+    if (wsAny._parkedKeepalive) {
+      clearInterval(wsAny._parkedKeepalive);
+      delete wsAny._parkedKeepalive;
+    }
     try {
       slot.openaiRealtime.close();
     } catch {
@@ -1014,7 +1033,7 @@ var Patter = class {
   }
   /** Run the agent in interactive terminal-test mode (no real telephony). */
   async test(opts) {
-    const { TestSession: TestSession2 } = await import("./test-mode-WEKKNBLD.mjs");
+    const { TestSession: TestSession2 } = await import("./test-mode-RS57BDM6.mjs");
     const session = new TestSession2();
     await session.run({
       agent: opts.agent,
@@ -1144,7 +1163,9 @@ var Patter = class {
     const tts = agent.tts;
     const sttOpen = typeof stt?.openParkedConnection === "function" ? stt.openParkedConnection.bind(stt) : null;
     const ttsOpen = typeof tts?.openParkedConnection === "function" ? tts.openParkedConnection.bind(tts) : null;
-    if (!sttOpen && !ttsOpen) return;
+    const providerStr = agent.provider ?? "";
+    const wantsRealtimePark = providerStr === "openai_realtime" || providerStr === "openai_realtime_2";
+    if (!sttOpen && !ttsOpen && !wantsRealtimePark) return;
     const slot = {};
     this.prewarmedConnections.set(callId, slot);
     const startedAt = Date.now();
@@ -1189,6 +1210,43 @@ var Patter = class {
         }
       })());
     }
+    if (wantsRealtimePark) {
+      tasks.push((async () => {
+        const { OpenAIRealtime2Adapter: OpenAIRealtime2Adapter2 } = await import("./openai-realtime-2-CNFARP25.mjs");
+        const apiKey = process.env.OPENAI_API_KEY ?? "";
+        if (!apiKey) {
+          getLogger().debug(`Park OpenAI Realtime skipped for ${callId}: no OPENAI_API_KEY`);
+          return;
+        }
+        try {
+          const tmpAdapter = new OpenAIRealtime2Adapter2(
+            apiKey,
+            agent.model ?? "gpt-realtime-mini",
+            agent.voice ?? "alloy",
+            agent.systemPrompt ?? "",
+            [],
+            // audioFormat — the GA adapter always emits audio/pcm@24000
+            // internally regardless of this value, but it's a required
+            // positional param. Default to g711_ulaw (Twilio wire format).
+            void 0
+          );
+          const ws = await tmpAdapter.openParkedConnection();
+          if (this.prewarmedConnections.get(callId) !== slot) {
+            try {
+              ws.close();
+            } catch {
+            }
+            return;
+          }
+          slot.openaiRealtime = ws;
+          getLogger().info(
+            `[PREWARM] callId=${callId} provider=openai_realtime ms=${Date.now() - startedAt}`
+          );
+        } catch (err) {
+          getLogger().debug(`Park OpenAI Realtime failed for ${callId}: ${String(err)}`);
+        }
+      })());
+    }
     const task = (async () => {
       await Promise.allSettled(tasks);
     })();
@@ -1266,7 +1324,7 @@ var Patter = class {
    * with a warn when the cap is reached (the call still proceeds —
    * StreamHandler falls back to live TTS).
    */
-  spawnPrewarmFirstMessage(agent, callId, ringTimeout) {
+  spawnPrewarmFirstMessage(agent, callId, ringTimeout, carrier) {
     if (!agent.prewarmFirstMessage) return;
     const providerMode = agent.provider ?? "openai_realtime";
     if (providerMode !== "pipeline") {
@@ -1279,6 +1337,18 @@ var Patter = class {
     const tts = agent.tts;
     if (!firstMessage || !tts) return;
     if (typeof tts.synthesizeStream !== "function") return;
+    if (carrier) {
+      const carrierAware = tts;
+      if (typeof carrierAware.setTelephonyCarrier === "function") {
+        try {
+          carrierAware.setTelephonyCarrier(carrier);
+        } catch (err) {
+          getLogger().debug(
+            `Prewarm TTS setTelephonyCarrier failed for ${callId}: ${String(err)}`
+          );
+        }
+      }
+    }
     const inFlight = this.prewarmAudio.size + this.prewarmTasks.size;
     if (inFlight >= PREWARM_CACHE_MAX) {
       getLogger().warn(
@@ -1391,16 +1461,25 @@ var Patter = class {
         telnyxCallId = body.data?.call_control_id;
       } catch {
       }
-      if (this.embeddedServer && telnyxCallId) {
-        this.embeddedServer.metricsStore.recordCallInitiated({
+      if (telnyxCallId) {
+        const initiatedPayload = {
           call_id: telnyxCallId,
           caller: phoneNumber,
           callee: options.to,
-          direction: "outbound"
-        });
+          direction: "outbound",
+          status: "initiated"
+        };
+        if (this.embeddedServer) {
+          this.embeddedServer.metricsStore.recordCallInitiated(initiatedPayload);
+        }
+        try {
+          const { notifyDashboard: notifyDashboard2 } = await import("./persistence-LVIAHESK.mjs");
+          notifyDashboard2(initiatedPayload);
+        } catch {
+        }
       }
       if (telnyxCallId) {
-        this.spawnPrewarmFirstMessage(options.agent, telnyxCallId, effectiveRingTimeout);
+        this.spawnPrewarmFirstMessage(options.agent, telnyxCallId, effectiveRingTimeout, "telnyx");
         if (options.agent.prewarm !== false) {
           this.parkProviderConnections(options.agent, telnyxCallId);
         }
@@ -1453,21 +1532,30 @@ var Patter = class {
       twilioNotificationsPath = body.subresource_uris?.notifications;
     } catch {
     }
-    if (this.embeddedServer && twilioCallSid) {
-      this.embeddedServer.metricsStore.recordCallInitiated({
+    if (twilioCallSid) {
+      const initiatedPayload = {
         call_id: twilioCallSid,
         caller: phoneNumber,
         callee: options.to,
-        direction: "outbound"
-      });
-      if (twilioNotificationsPath) {
-        getLogger().info(
-          `Outbound call ${twilioCallSid} placed. Twilio notifications: https://api.twilio.com${twilioNotificationsPath} (check here if the call drops with no audio).`
-        );
+        direction: "outbound",
+        status: "initiated"
+      };
+      if (this.embeddedServer) {
+        this.embeddedServer.metricsStore.recordCallInitiated(initiatedPayload);
+        if (twilioNotificationsPath) {
+          getLogger().info(
+            `Outbound call ${twilioCallSid} placed. Twilio notifications: https://api.twilio.com${twilioNotificationsPath} (check here if the call drops with no audio).`
+          );
+        }
+      }
+      try {
+        const { notifyDashboard: notifyDashboard2 } = await import("./persistence-LVIAHESK.mjs");
+        notifyDashboard2(initiatedPayload);
+      } catch {
       }
     }
     if (twilioCallSid) {
-      this.spawnPrewarmFirstMessage(options.agent, twilioCallSid, effectiveRingTimeout);
+      this.spawnPrewarmFirstMessage(options.agent, twilioCallSid, effectiveRingTimeout, "twilio");
       if (options.agent.prewarm !== false) {
         this.parkProviderConnections(options.agent, twilioCallSid);
       }
@@ -2764,109 +2852,694 @@ function scheduleInterval(intervalOrOpts, callback) {
   };
 }
 
-// src/stt/deepgram.ts
+// src/providers/elevenlabs-tts.ts
 init_esm_shims();
-var STT = class extends DeepgramSTT {
-  static providerKey = "deepgram";
-  constructor(opts = {}) {
-    const key = opts.apiKey ?? process.env.DEEPGRAM_API_KEY;
-    if (!key) {
-      throw new Error(
-        "Deepgram STT requires an apiKey. Pass { apiKey: 'dg_...' } or set DEEPGRAM_API_KEY in the environment."
-      );
-    }
-    super(
-      key,
-      opts.language ?? "en",
-      opts.model ?? "nova-3",
-      opts.encoding ?? "linear16",
-      opts.sampleRate ?? 16e3,
-      {
-        endpointingMs: opts.endpointingMs ?? 150,
-        utteranceEndMs: opts.utteranceEndMs === null ? null : opts.utteranceEndMs ?? 1e3,
-        smartFormat: opts.smartFormat ?? true,
-        interimResults: opts.interimResults ?? true,
-        ...opts.vadEvents !== void 0 ? { vadEvents: opts.vadEvents } : {}
-      }
-    );
-  }
+var ELEVENLABS_BASE_URL = "https://api.elevenlabs.io/v1";
+var ELEVENLABS_VOICE_ID_BY_NAME = {
+  rachel: "21m00Tcm4TlvDq8ikWAM",
+  drew: "29vD33N1CtxCmqQRPOHJ",
+  clyde: "2EiwWnXFnvU5JabPnv8n",
+  paul: "5Q0t7uMcjvnagumLfvZi",
+  domi: "AZnzlk1XvdvUeBnXmlld",
+  dave: "CYw3kZ02Hs0563khs1Fj",
+  fin: "D38z5RcWu1voky8WS1ja",
+  bella: "EXAVITQu4vr4xnSDxMaL",
+  antoni: "ErXwobaYiN019PkySvjV",
+  thomas: "GBv7mTt0atIp3Br8iCZE",
+  charlie: "IKne3meq5aSn9XLyUdCD",
+  george: "JBFqnCBsd6RMkjVDRZzb",
+  emily: "LcfcDJNUP1GQjkzn1xUU",
+  elli: "MF3mGyEYCl7XYWbV9V6O",
+  callum: "N2lVS1w4EtoT3dr4eOWO",
+  patrick: "ODq5zmih8GrVes37Dizd",
+  harry: "SOYHLrjzK2X1ezoPC6cr",
+  liam: "TX3LPaxmHKxFdv7VOQHJ",
+  dorothy: "ThT5KcBeYPX3keUQqHPh",
+  josh: "TxGEqnHWrfWFTfGW9XjX",
+  arnold: "VR6AewLTigWG4xSOukaG",
+  charlotte: "XB0fDUnXU5powFXDhCwa",
+  matilda: "XrExE9yKIg1WjnnlVkGX",
+  matthew: "Yko7PKHZNXotIFUBG7I9",
+  james: "ZQe5CZNOzWyzPSCn5a3c",
+  joseph: "Zlb1dXrM653N07WRdFW3",
+  jeremy: "bVMeCyTHy58xNoL34h3p",
+  michael: "flq6f7yk4E4fJM5XTYuZ",
+  ethan: "g5CIjZEefAph4nQFvHAz",
+  gigi: "jBpfuIE2acCO8z3wKNLl",
+  freya: "jsCqWAovK2LkecY7zXl4",
+  brian: "nPczCjzI2devNBz1zQrb",
+  grace: "oWAxZDx7w5VEj9dCyTzz",
+  daniel: "onwK4e9ZLuTAKqWW03F9",
+  lily: "pFZP5JQG7iQjIQuC4Bku",
+  serena: "pMsXgVXv3BLzUgSXRplE",
+  adam: "pNInz6obpgDQGcFmaJgB",
+  nicole: "piTKgcLEGmPE4e6mEKli",
+  bill: "pqHfZKP75CvOlQylNhV4",
+  jessie: "t0jbNlBVZ17f02VDIeMI",
+  ryan: "wViXBPUzp2ZZixB1xQuM",
+  sam: "yoZ06aMxZJJ28mfd3POQ",
+  glinda: "z9fAnlkpzviPz146aGWa",
+  giovanni: "zcAOhNBS3c14rBihAFp1",
+  mimi: "zrHiDhphv9ZnVXBqCLjz",
+  sarah: "EXAVITQu4vr4xnSDxMaL",
+  alloy: "EXAVITQu4vr4xnSDxMaL"
 };
-
-// src/stt/whisper.ts
-init_esm_shims();
-
-// src/providers/whisper-stt.ts
-init_esm_shims();
-var OPENAI_TRANSCRIPTION_URL = "https://api.openai.com/v1/audio/transcriptions";
-var DEFAULT_BUFFER_SIZE = 16e3 * 2;
-var ALLOWED_MODELS = /* @__PURE__ */ new Set(["whisper-1", "gpt-4o-transcribe", "gpt-4o-mini-transcribe"]);
-function wrapPcmInWav(pcm, sampleRate = 16e3, channels = 1, bitsPerSample = 16) {
-  const dataSize = pcm.length;
-  const header = Buffer.alloc(44);
-  header.write("RIFF", 0);
-  header.writeUInt32LE(36 + dataSize, 4);
-  header.write("WAVE", 8);
-  header.write("fmt ", 12);
-  header.writeUInt32LE(16, 16);
-  header.writeUInt16LE(1, 20);
-  header.writeUInt16LE(channels, 22);
-  header.writeUInt32LE(sampleRate, 24);
-  header.writeUInt32LE(sampleRate * channels * (bitsPerSample / 8), 28);
-  header.writeUInt16LE(channels * (bitsPerSample / 8), 32);
-  header.writeUInt16LE(bitsPerSample, 34);
-  header.write("data", 36);
-  header.writeUInt32LE(dataSize, 40);
-  return Buffer.concat([header, pcm]);
+var VOICE_ID_PATTERN = /^[A-Za-z0-9]{20}$/;
+function resolveVoiceId(voice) {
+  if (!voice) return voice;
+  if (VOICE_ID_PATTERN.test(voice)) return voice;
+  return ELEVENLABS_VOICE_ID_BY_NAME[voice.toLowerCase()] ?? voice;
 }
-var WhisperSTT = class _WhisperSTT {
-  /** Stable pricing/dashboard key — read by stream-handler/metrics. */
-  static providerKey = "whisper";
+var ElevenLabsModel = {
+  V3: "eleven_v3",
+  FLASH_V2_5: "eleven_flash_v2_5",
+  TURBO_V2_5: "eleven_turbo_v2_5",
+  MULTILINGUAL_V2: "eleven_multilingual_v2",
+  MONOLINGUAL_V1: "eleven_monolingual_v1"
+};
+var ElevenLabsOutputFormat = {
+  MP3_22050_32: "mp3_22050_32",
+  MP3_44100_32: "mp3_44100_32",
+  MP3_44100_64: "mp3_44100_64",
+  MP3_44100_96: "mp3_44100_96",
+  MP3_44100_128: "mp3_44100_128",
+  MP3_44100_192: "mp3_44100_192",
+  PCM_8000: "pcm_8000",
+  PCM_16000: "pcm_16000",
+  PCM_22050: "pcm_22050",
+  PCM_24000: "pcm_24000",
+  PCM_44100: "pcm_44100",
+  ULAW_8000: "ulaw_8000"
+};
+var ElevenLabsTTS = class _ElevenLabsTTS {
+  // Stable pricing/dashboard key — read by stream-handler / metrics via
+  // ``(agent.tts.constructor as any).providerKey``. Without this the cost
+  // calculator falls back to ``constructor.name`` ("ElevenLabsTTS") which
+  // does NOT match the pricing table key "elevenlabs", silently zeroing
+  // TTS cost for callers that construct the raw REST class directly
+  // (exposed at top level as ``ElevenLabsRestTTS``).
+  static providerKey = "elevenlabs";
   apiKey;
-  model;
-  language;
-  bufferSize;
-  responseFormat;
-  // Accumulate chunks in an array and concat once on flush — avoids the
-  // per-``sendAudio`` O(n) ``Buffer.concat([buffer, chunk])`` that quickly
-  // dominates CPU when the phone leg delivers 20 ms frames.
-  chunks = [];
-  bufferedBytes = 0;
-  callbacks = /* @__PURE__ */ new Set();
-  running = false;
-  pendingTranscriptions = [];
+  voiceId;
+  modelId;
+  _outputFormat;
+  _outputFormatExplicit;
+  voiceSettings;
+  languageCode;
+  chunkSize;
   /**
-   * @param apiKey OpenAI API key.
-   * @param language ISO-639-1 language code (e.g. ``"en"``, ``"it"``). Optional.
-   * @param model One of ``whisper-1``, ``gpt-4o-transcribe``, ``gpt-4o-mini-transcribe``.
-   * @param bufferSize Bytes of PCM16 to buffer before each transcription request.
-   * @param responseFormat ``"json"`` (default) or ``"verbose_json"``.
-   *
-   * Argument order matches the Python SDK's ``WhisperSTT(api_key, language, model, response_format)``
-   * for cross-language parity. Pre-0.5.3 the TS positional order was
-   * ``(apiKey, model, language, bufferSize, responseFormat)`` — callers using
-   * the old order will need to swap ``language`` and ``model``.
+   * Public view of the (possibly auto-flipped) wire format. Read by the
+   * stream-handler to decide whether to skip the client-side resample +
+   * mulaw encode when the bytes are already in the carrier's wire codec.
    */
-  constructor(apiKey, language, model = "whisper-1", bufferSize = DEFAULT_BUFFER_SIZE, responseFormat = "json") {
-    if (!ALLOWED_MODELS.has(model)) {
-      throw new Error(
-        `WhisperSTT: unsupported model "${model}". Expected one of ${[...ALLOWED_MODELS].join(", ")}.`
-      );
-    }
-    this.apiKey = apiKey;
-    this.model = model;
-    this.language = language;
-    this.bufferSize = bufferSize;
-    this.responseFormat = responseFormat;
-  }
-  /** Factory for Twilio calls — mulaw 8 kHz is transcoded upstream, so we still receive PCM 16-bit. */
-  static forTwilio(apiKey, language = "en", model = "whisper-1") {
-    return new _WhisperSTT(apiKey, language, model);
-  }
-  /** Reset the audio buffer and arm the adapter for incoming chunks. */
-  async connect() {
-    this.running = true;
-    this.chunks = [];
-    this.bufferedBytes = 0;
+  get outputFormat() {
+    return this._outputFormat;
+  }
+  constructor(apiKey, voiceIdOrOptions = "21m00Tcm4TlvDq8ikWAM", modelId = ElevenLabsModel.FLASH_V2_5, outputFormat = ElevenLabsOutputFormat.PCM_16000) {
+    this.apiKey = apiKey;
+    if (typeof voiceIdOrOptions === "object") {
+      const o = voiceIdOrOptions;
+      this.voiceId = resolveVoiceId(o.voiceId ?? "21m00Tcm4TlvDq8ikWAM");
+      this.modelId = o.modelId ?? ElevenLabsModel.FLASH_V2_5;
+      this._outputFormatExplicit = o.outputFormat !== void 0;
+      this._outputFormat = o.outputFormat ?? ElevenLabsOutputFormat.PCM_16000;
+      this.voiceSettings = o.voiceSettings;
+      this.languageCode = o.languageCode;
+      this.chunkSize = o.chunkSize ?? 4096;
+    } else {
+      this.voiceId = resolveVoiceId(voiceIdOrOptions);
+      this.modelId = modelId;
+      this._outputFormatExplicit = outputFormat !== ElevenLabsOutputFormat.PCM_16000;
+      this._outputFormat = outputFormat;
+      this.voiceSettings = void 0;
+      this.languageCode = void 0;
+      this.chunkSize = 4096;
+    }
+  }
+  /**
+   * Hook called by ``StreamHandler.initPipeline`` to advise the carrier
+   * wire format. When the user did NOT pass an explicit ``outputFormat``,
+   * auto-flip to the carrier's native codec so the audio bytes ElevenLabs
+   * returns are already in Twilio/Telnyx wire format — eliminating the
+   * client-side 16 kHz → 8 kHz resample and PCM → μ-law encode. The
+   * resample/encode chain was a source of audible artifacts on the
+   * prewarmed firstMessage (see 0.6.2 acceptance notes — burst delivery
+   * of resampled audio crackled on the carrier-side jitter buffer).
+   *
+   * No-op when the caller passed an explicit ``outputFormat`` (incl. via
+   * the ``forTwilio`` / ``forTelnyx`` factories) — user wins.
+   *
+   * Parity with {@link ElevenLabsWebSocketTTS.setTelephonyCarrier}.
+   */
+  setTelephonyCarrier(carrier) {
+    if (this._outputFormatExplicit) return;
+    if (carrier === "twilio") {
+      this._outputFormat = ElevenLabsOutputFormat.ULAW_8000;
+    } else if (carrier === "telnyx") {
+      this._outputFormat = ElevenLabsOutputFormat.PCM_16000;
+    }
+  }
+  /**
+   * Construct an instance pre-configured for Twilio Media Streams.
+   *
+   * Sets `outputFormat='ulaw_8000'` so ElevenLabs emits μ-law @ 8 kHz
+   * directly — the exact wire format Twilio's media stream uses — letting
+   * the SDK skip the 16 kHz→8 kHz resample and PCM→μ-law conversion in
+   * `TwilioAudioSender`. Saves ~30–80 ms first-byte and per-frame CPU,
+   * and removes a potential aliasing source.
+   *
+   * `voiceSettings` defaults to a low-bandwidth-friendly profile
+   * (speaker boost off, modest stability) which sounds cleaner at 8 kHz
+   * μ-law than the studio default. Pass an explicit object to override.
+   */
+  static forTwilio(apiKey, options = {}) {
+    const voiceSettings = options.voiceSettings ?? {
+      // Speaker boost adds high-frequency emphasis that aliases ugly over an
+      // 8 kHz μ-law line. Slightly higher stability tames the excursions
+      // that compander quantization noise can amplify.
+      stability: 0.6,
+      similarity_boost: 0.75,
+      use_speaker_boost: false
+    };
+    return new _ElevenLabsTTS(apiKey, {
+      ...options,
+      voiceSettings,
+      outputFormat: ElevenLabsOutputFormat.ULAW_8000
+    });
+  }
+  /**
+   * Construct an instance pre-configured for Telnyx bidirectional media.
+   *
+   * Telnyx's default media-streaming codec is L16 PCM @ 16 kHz, which
+   * matches our default Telnyx handler. We pick `pcm_16000` so the audio
+   * flows end-to-end with zero resampling or transcoding.
+   *
+   * Trade-off: if your Telnyx profile is pinned to PCMU/8000 (μ-law),
+   * construct `ElevenLabsTTS` directly with `outputFormat: 'ulaw_8000'`
+   * — Telnyx supports that natively too.
+   */
+  static forTelnyx(apiKey, options = {}) {
+    return new _ElevenLabsTTS(apiKey, {
+      ...options,
+      outputFormat: ElevenLabsOutputFormat.PCM_16000
+    });
+  }
+  /**
+   * Synthesise text to speech and return the full audio as a single Buffer.
+   *
+   * For large chunks (or when latency matters) call `synthesizeStream` instead.
+   */
+  async synthesize(text) {
+    const chunks = [];
+    for await (const chunk of this.synthesizeStream(text)) {
+      chunks.push(chunk);
+    }
+    return Buffer.concat(chunks);
+  }
+  /**
+   * Synthesise text and yield audio chunks as they arrive (streaming).
+   *
+   * The yielded buffers are raw PCM at 16 kHz (or whatever `outputFormat` is
+   * configured to). `chunkSize` controls the maximum yield size — 512 is a
+   * good choice for low-latency telephony.
+   */
+  async *synthesizeStream(text) {
+    const url = `${ELEVENLABS_BASE_URL}/text-to-speech/${encodeURIComponent(this.voiceId)}/stream?output_format=${encodeURIComponent(this._outputFormat)}`;
+    const body = {
+      text,
+      model_id: this.modelId
+    };
+    if (this.voiceSettings) body["voice_settings"] = this.voiceSettings;
+    if (this.languageCode) body["language_code"] = this.languageCode;
+    const response = await fetch(url, {
+      method: "POST",
+      headers: {
+        "xi-api-key": this.apiKey,
+        "Content-Type": "application/json"
+      },
+      body: JSON.stringify(body),
+      signal: AbortSignal.timeout(3e4)
+    });
+    if (!response.ok) {
+      const errBody = await response.text();
+      throw new Error(`ElevenLabs TTS error ${response.status}: ${errBody}`);
+    }
+    if (!response.body) {
+      throw new Error("ElevenLabs TTS: no response body");
+    }
+    const reader = response.body.getReader();
+    try {
+      while (true) {
+        const { done, value } = await reader.read();
+        if (done) break;
+        if (!value || value.length === 0) continue;
+        const buf = Buffer.from(value);
+        for (let offset = 0; offset < buf.length; offset += this.chunkSize) {
+          yield buf.subarray(offset, Math.min(offset + this.chunkSize, buf.length));
+        }
+      }
+    } finally {
+      if (typeof reader.cancel === "function") await reader.cancel().catch(() => {
+      });
+      reader.releaseLock();
+    }
+  }
+};
+
+// src/providers/cartesia-tts.ts
+init_esm_shims();
+var CARTESIA_BASE_URL = "https://api.cartesia.ai";
+var CARTESIA_API_VERSION = "2025-04-16";
+var CARTESIA_DEFAULT_VOICE_ID = "f786b574-daa5-4673-aa0c-cbe3e8534c02";
+var CartesiaTTSModel = {
+  SONIC_3: "sonic-3",
+  SONIC_2: "sonic-2",
+  SONIC: "sonic"
+};
+var CartesiaTTSContainer = {
+  RAW: "raw",
+  WAV: "wav",
+  MP3: "mp3"
+};
+var CartesiaTTSEncoding = {
+  PCM_S16LE: "pcm_s16le",
+  PCM_F32LE: "pcm_f32le",
+  PCM_MULAW: "pcm_mulaw",
+  PCM_ALAW: "pcm_alaw"
+};
+var CartesiaTTSSampleRate = {
+  HZ_8000: 8e3,
+  HZ_16000: 16e3,
+  HZ_22050: 22050,
+  HZ_24000: 24e3,
+  HZ_44100: 44100
+};
+var CartesiaTTSVoiceMode = {
+  ID: "id",
+  EMBEDDING: "embedding"
+};
+var CartesiaTTS = class _CartesiaTTS {
+  /** Stable pricing/dashboard key — read by stream-handler/metrics. */
+  static providerKey = "cartesia_tts";
+  apiKey;
+  model;
+  voice;
+  language;
+  sampleRate;
+  speed;
+  emotion;
+  volume;
+  baseUrl;
+  apiVersion;
+  constructor(apiKey, opts = {}) {
+    this.apiKey = apiKey;
+    this.model = opts.model ?? CartesiaTTSModel.SONIC_3;
+    this.voice = opts.voice ?? CARTESIA_DEFAULT_VOICE_ID;
+    this.language = opts.language ?? "en";
+    this.sampleRate = opts.sampleRate ?? CartesiaTTSSampleRate.HZ_16000;
+    this.speed = opts.speed;
+    this.emotion = typeof opts.emotion === "string" ? [opts.emotion] : opts.emotion;
+    this.volume = opts.volume;
+    this.baseUrl = opts.baseUrl ?? CARTESIA_BASE_URL;
+    this.apiVersion = opts.apiVersion ?? CARTESIA_API_VERSION;
+  }
+  /**
+   * Construct an instance pre-configured for Twilio Media Streams.
+   *
+   * Sets `sampleRate=8000` so Cartesia emits PCM_S16LE @ 8 kHz directly.
+   * Twilio's media stream uses μ-law @ 8 kHz so the SDK still does the
+   * PCM → μ-law transcode client-side, but the 16 kHz → 8 kHz resample
+   * step is skipped. Saves ~10–30 ms first-byte plus per-frame CPU and
+   * removes a potential aliasing source.
+   */
+  static forTwilio(apiKey, options = {}) {
+    return new _CartesiaTTS(apiKey, {
+      ...options,
+      sampleRate: CartesiaTTSSampleRate.HZ_8000
+    });
+  }
+  /**
+   * Construct an instance pre-configured for Telnyx bidirectional media.
+   *
+   * Sets `sampleRate=16000` to match Telnyx's L16/16000 default codec —
+   * audio flows end-to-end with zero resampling or transcoding. Same as
+   * the bare-constructor default; exists for API symmetry with
+   * {@link CartesiaTTS.forTwilio}.
+   */
+  static forTelnyx(apiKey, options = {}) {
+    return new _CartesiaTTS(apiKey, {
+      ...options,
+      sampleRate: CartesiaTTSSampleRate.HZ_16000
+    });
+  }
+  /** Build the JSON payload for the Cartesia bytes endpoint. */
+  buildPayload(text) {
+    const payload = {
+      model_id: this.model,
+      voice: { mode: CartesiaTTSVoiceMode.ID, id: this.voice },
+      transcript: text,
+      output_format: {
+        container: CartesiaTTSContainer.RAW,
+        encoding: CartesiaTTSEncoding.PCM_S16LE,
+        sample_rate: this.sampleRate
+      },
+      language: this.language
+    };
+    const generationConfig = {};
+    if (this.speed !== void 0) generationConfig.speed = this.speed;
+    if (this.emotion && this.emotion.length > 0)
+      generationConfig.emotion = this.emotion[0];
+    if (this.volume !== void 0) generationConfig.volume = this.volume;
+    if (Object.keys(generationConfig).length > 0) {
+      payload.generation_config = generationConfig;
+    }
+    return payload;
+  }
+  /**
+   * Pre-call HTTP warmup for the Cartesia `/tts/bytes` endpoint.
+   *
+   * Issues a lightweight `GET <baseUrl>/voices` so DNS, TLS, and HTTP/2
+   * are already up by the time the first `synthesizeStream()` POST
+   * lands. Best-effort: 5 s timeout, all exceptions swallowed at
+   * debug level.
+   *
+   * Billing safety: `GET /voices` is a free metadata read on
+   * Cartesia's REST surface (per https://docs.cartesia.ai). It does
+   * not consume synthesis credits. The actual synthesis is billed
+   * only when `POST /tts/bytes` runs with a non-empty `transcript`.
+   *
+   * Note: Cartesia TTS uses the HTTP path (vs the WebSocket variant
+   * Cartesia also exposes) — connection warmup is therefore HTTP-GET
+   * based, not WebSocket pre-handshake. The latency win is smaller
+   * (~50-150 ms vs the ~200-500 ms of a WS prewarm) but still real.
+   */
+  async warmup() {
+    try {
+      await fetch(`${this.baseUrl}/voices`, {
+        method: "GET",
+        headers: {
+          "X-API-Key": this.apiKey,
+          "Cartesia-Version": this.apiVersion
+        },
+        signal: AbortSignal.timeout(5e3)
+      });
+    } catch (err) {
+      getLogger().debug(`Cartesia TTS warmup failed (best-effort): ${String(err)}`);
+    }
+  }
+  /** Synthesize text and return the concatenated audio buffer. */
+  async synthesize(text) {
+    const chunks = [];
+    for await (const chunk of this.synthesizeStream(text)) {
+      chunks.push(chunk);
+    }
+    return Buffer.concat(chunks);
+  }
+  /**
+   * Synthesize text and yield raw PCM_S16LE chunks at the configured
+   * `sampleRate` as they arrive from Cartesia.
+   */
+  async *synthesizeStream(text) {
+    const response = await fetch(`${this.baseUrl}/tts/bytes`, {
+      method: "POST",
+      headers: {
+        "X-API-Key": this.apiKey,
+        "Cartesia-Version": this.apiVersion,
+        "Content-Type": "application/json"
+      },
+      body: JSON.stringify(this.buildPayload(text)),
+      signal: AbortSignal.timeout(3e4)
+    });
+    if (!response.ok) {
+      const body = await response.text();
+      throw new Error(`Cartesia TTS error ${response.status}: ${body}`);
+    }
+    if (!response.body) {
+      throw new Error("Cartesia TTS: no response body");
+    }
+    const reader = response.body.getReader();
+    try {
+      while (true) {
+        const { done, value } = await reader.read();
+        if (done) break;
+        if (value && value.length > 0) {
+          yield Buffer.from(value);
+        }
+      }
+    } finally {
+      if (typeof reader.cancel === "function")
+        await reader.cancel().catch(() => {
+        });
+      reader.releaseLock();
+    }
+  }
+};
+
+// src/providers/rime-tts.ts
+init_esm_shims();
+var RIME_BASE_URL = "https://users.rime.ai/v1/rime-tts";
+var RimeModel = {
+  ARCANA: "arcana",
+  MIST: "mist",
+  MIST_V2: "mistv2"
+};
+var RimeAudioFormat = {
+  PCM: "audio/pcm",
+  MP3: "audio/mp3",
+  WAV: "audio/wav",
+  MULAW: "audio/mulaw"
+};
+var ARCANA_MODEL_TIMEOUT_MS = 60 * 4 * 1e3;
+var MIST_MODEL_TIMEOUT_MS = 30 * 1e3;
+function isMistModel(model) {
+  return model.includes(RimeModel.MIST);
+}
+function timeoutForModel(model) {
+  if (model === RimeModel.ARCANA) return ARCANA_MODEL_TIMEOUT_MS;
+  return MIST_MODEL_TIMEOUT_MS;
+}
+var RimeTTS = class {
+  /** Stable pricing/dashboard key — read by stream-handler/metrics. */
+  static providerKey = "rime";
+  apiKey;
+  model;
+  speaker;
+  lang;
+  sampleRate;
+  repetitionPenalty;
+  temperature;
+  topP;
+  maxTokens;
+  speedAlpha;
+  reduceLatency;
+  pauseBetweenBrackets;
+  phonemizeBetweenBrackets;
+  baseUrl;
+  totalTimeoutMs;
+  constructor(apiKey, opts = {}) {
+    this.apiKey = apiKey;
+    this.model = opts.model ?? RimeModel.ARCANA;
+    const defaultSpeaker = isMistModel(this.model) ? "cove" : "astra";
+    this.speaker = opts.speaker ?? defaultSpeaker;
+    this.lang = opts.lang ?? "eng";
+    this.sampleRate = opts.sampleRate ?? 16e3;
+    this.repetitionPenalty = opts.repetitionPenalty;
+    this.temperature = opts.temperature;
+    this.topP = opts.topP;
+    this.maxTokens = opts.maxTokens;
+    this.speedAlpha = opts.speedAlpha;
+    this.reduceLatency = opts.reduceLatency;
+    this.pauseBetweenBrackets = opts.pauseBetweenBrackets;
+    this.phonemizeBetweenBrackets = opts.phonemizeBetweenBrackets;
+    this.baseUrl = opts.baseUrl ?? RIME_BASE_URL;
+    this.totalTimeoutMs = timeoutForModel(this.model);
+  }
+  buildPayload(text) {
+    const payload = {
+      speaker: this.speaker,
+      text,
+      modelId: this.model
+    };
+    if (this.model === RimeModel.ARCANA) {
+      if (this.repetitionPenalty !== void 0)
+        payload.repetition_penalty = this.repetitionPenalty;
+      if (this.temperature !== void 0) payload.temperature = this.temperature;
+      if (this.topP !== void 0) payload.top_p = this.topP;
+      if (this.maxTokens !== void 0) payload.max_tokens = this.maxTokens;
+      payload.lang = this.lang;
+      payload.samplingRate = this.sampleRate;
+    } else if (isMistModel(this.model)) {
+      payload.lang = this.lang;
+      payload.samplingRate = this.sampleRate;
+      if (this.speedAlpha !== void 0) payload.speedAlpha = this.speedAlpha;
+      if (this.model === RimeModel.MIST_V2 && this.reduceLatency !== void 0) {
+        payload.reduceLatency = this.reduceLatency;
+      }
+      if (this.pauseBetweenBrackets !== void 0) {
+        payload.pauseBetweenBrackets = this.pauseBetweenBrackets;
+      }
+      if (this.phonemizeBetweenBrackets !== void 0) {
+        payload.phonemizeBetweenBrackets = this.phonemizeBetweenBrackets;
+      }
+    }
+    return payload;
+  }
+  /** Synthesize text and return the concatenated audio buffer. */
+  async synthesize(text) {
+    const chunks = [];
+    for await (const chunk of this.synthesizeStream(text)) {
+      chunks.push(chunk);
+    }
+    return Buffer.concat(chunks);
+  }
+  /**
+   * Synthesize text and yield raw PCM_S16LE chunks at the configured
+   * `sampleRate` as they stream in.
+   */
+  async *synthesizeStream(text) {
+    const response = await fetch(this.baseUrl, {
+      method: "POST",
+      headers: {
+        accept: RimeAudioFormat.PCM,
+        Authorization: `Bearer ${this.apiKey}`,
+        "content-type": "application/json"
+      },
+      body: JSON.stringify(this.buildPayload(text)),
+      signal: AbortSignal.timeout(this.totalTimeoutMs)
+    });
+    if (!response.ok) {
+      const body = await response.text();
+      throw new Error(`Rime TTS error ${response.status}: ${body}`);
+    }
+    const contentType = response.headers.get("content-type") ?? "";
+    if (!contentType.startsWith("audio")) {
+      const body = await response.text();
+      throw new Error(`Rime returned non-audio response: ${body.slice(0, 500)}`);
+    }
+    if (!response.body) {
+      throw new Error("Rime TTS: no response body");
+    }
+    const reader = response.body.getReader();
+    try {
+      while (true) {
+        const { done, value } = await reader.read();
+        if (done) break;
+        if (value && value.length > 0) {
+          yield Buffer.from(value);
+        }
+      }
+    } finally {
+      if (typeof reader.cancel === "function")
+        await reader.cancel().catch(() => {
+        });
+      reader.releaseLock();
+    }
+  }
+};
+
+// src/stt/deepgram.ts
+init_esm_shims();
+var STT = class extends DeepgramSTT {
+  static providerKey = "deepgram";
+  constructor(opts = {}) {
+    const key = opts.apiKey ?? process.env.DEEPGRAM_API_KEY;
+    if (!key) {
+      throw new Error(
+        "Deepgram STT requires an apiKey. Pass { apiKey: 'dg_...' } or set DEEPGRAM_API_KEY in the environment."
+      );
+    }
+    super(
+      key,
+      opts.language ?? "en",
+      opts.model ?? "nova-3",
+      opts.encoding ?? "linear16",
+      opts.sampleRate ?? 16e3,
+      {
+        endpointingMs: opts.endpointingMs ?? 150,
+        utteranceEndMs: opts.utteranceEndMs === null ? null : opts.utteranceEndMs ?? 1e3,
+        smartFormat: opts.smartFormat ?? true,
+        interimResults: opts.interimResults ?? true,
+        ...opts.vadEvents !== void 0 ? { vadEvents: opts.vadEvents } : {}
+      }
+    );
+  }
+};
+
+// src/stt/whisper.ts
+init_esm_shims();
+
+// src/providers/whisper-stt.ts
+init_esm_shims();
+var OPENAI_TRANSCRIPTION_URL = "https://api.openai.com/v1/audio/transcriptions";
+var DEFAULT_BUFFER_SIZE = 16e3 * 2;
+var ALLOWED_MODELS = /* @__PURE__ */ new Set(["whisper-1", "gpt-4o-transcribe", "gpt-4o-mini-transcribe"]);
+function wrapPcmInWav(pcm, sampleRate = 16e3, channels = 1, bitsPerSample = 16) {
+  const dataSize = pcm.length;
+  const header = Buffer.alloc(44);
+  header.write("RIFF", 0);
+  header.writeUInt32LE(36 + dataSize, 4);
+  header.write("WAVE", 8);
+  header.write("fmt ", 12);
+  header.writeUInt32LE(16, 16);
+  header.writeUInt16LE(1, 20);
+  header.writeUInt16LE(channels, 22);
+  header.writeUInt32LE(sampleRate, 24);
+  header.writeUInt32LE(sampleRate * channels * (bitsPerSample / 8), 28);
+  header.writeUInt16LE(channels * (bitsPerSample / 8), 32);
+  header.writeUInt16LE(bitsPerSample, 34);
+  header.write("data", 36);
+  header.writeUInt32LE(dataSize, 40);
+  return Buffer.concat([header, pcm]);
+}
+var WhisperSTT = class _WhisperSTT {
+  /** Stable pricing/dashboard key — read by stream-handler/metrics. */
+  static providerKey = "whisper";
+  apiKey;
+  model;
+  language;
+  bufferSize;
+  responseFormat;
+  // Accumulate chunks in an array and concat once on flush — avoids the
+  // per-``sendAudio`` O(n) ``Buffer.concat([buffer, chunk])`` that quickly
+  // dominates CPU when the phone leg delivers 20 ms frames.
+  chunks = [];
+  bufferedBytes = 0;
+  callbacks = /* @__PURE__ */ new Set();
+  running = false;
+  pendingTranscriptions = [];
+  /**
+   * @param apiKey OpenAI API key.
+   * @param language ISO-639-1 language code (e.g. ``"en"``, ``"it"``). Optional.
+   * @param model One of ``whisper-1``, ``gpt-4o-transcribe``, ``gpt-4o-mini-transcribe``.
+   * @param bufferSize Bytes of PCM16 to buffer before each transcription request.
+   * @param responseFormat ``"json"`` (default) or ``"verbose_json"``.
+   *
+   * Argument order matches the Python SDK's ``WhisperSTT(api_key, language, model, response_format)``
+   * for cross-language parity. Pre-0.5.3 the TS positional order was
+   * ``(apiKey, model, language, bufferSize, responseFormat)`` — callers using
+   * the old order will need to swap ``language`` and ``model``.
+   */
+  constructor(apiKey, language, model = "whisper-1", bufferSize = DEFAULT_BUFFER_SIZE, responseFormat = "json") {
+    if (!ALLOWED_MODELS.has(model)) {
+      throw new Error(
+        `WhisperSTT: unsupported model "${model}". Expected one of ${[...ALLOWED_MODELS].join(", ")}.`
+      );
+    }
+    this.apiKey = apiKey;
+    this.model = model;
+    this.language = language;
+    this.bufferSize = bufferSize;
+    this.responseFormat = responseFormat;
+  }
+  /** Factory for Twilio calls — mulaw 8 kHz is transcoded upstream, so we still receive PCM 16-bit. */
+  static forTwilio(apiKey, language = "en", model = "whisper-1") {
+    return new _WhisperSTT(apiKey, language, model);
+  }
+  /** Reset the audio buffer and arm the adapter for incoming chunks. */
+  async connect() {
+    this.running = true;
+    this.chunks = [];
+    this.bufferedBytes = 0;
   }
   /** Buffer a PCM16 chunk; flushes to Whisper once `bufferSize` bytes are reached. */
   sendAudio(audio) {
@@ -4448,264 +5121,42 @@ var SpeechmaticsSTT = class {
   close() {
     this.running = false;
     const ws = this.ws;
-    if (!ws) return;
-    this.ws = null;
-    const sendSafe = (payload) => {
-      if (ws.readyState === WebSocket5.OPEN) {
-        try {
-          ws.send(payload);
-        } catch {
-        }
-      }
-    };
-    sendSafe(
-      JSON.stringify({ message: "EndOfStream", last_seq_no: this.lastSeqNo })
-    );
-    try {
-      ws.close();
-    } catch {
-    }
-  }
-};
-
-// src/stt/speechmatics.ts
-var STT7 = class extends SpeechmaticsSTT {
-  static providerKey = "speechmatics";
-  constructor(opts = {}) {
-    const key = opts.apiKey ?? process.env.SPEECHMATICS_API_KEY;
-    if (!key) {
-      throw new Error(
-        "Speechmatics STT requires an apiKey. Pass { apiKey: 'sm_...' } or set SPEECHMATICS_API_KEY in the environment."
-      );
-    }
-    super(key, opts);
-  }
-};
-
-// src/tts/elevenlabs.ts
-init_esm_shims();
-
-// src/providers/elevenlabs-tts.ts
-init_esm_shims();
-var ELEVENLABS_BASE_URL = "https://api.elevenlabs.io/v1";
-var ELEVENLABS_VOICE_ID_BY_NAME = {
-  rachel: "21m00Tcm4TlvDq8ikWAM",
-  drew: "29vD33N1CtxCmqQRPOHJ",
-  clyde: "2EiwWnXFnvU5JabPnv8n",
-  paul: "5Q0t7uMcjvnagumLfvZi",
-  domi: "AZnzlk1XvdvUeBnXmlld",
-  dave: "CYw3kZ02Hs0563khs1Fj",
-  fin: "D38z5RcWu1voky8WS1ja",
-  bella: "EXAVITQu4vr4xnSDxMaL",
-  antoni: "ErXwobaYiN019PkySvjV",
-  thomas: "GBv7mTt0atIp3Br8iCZE",
-  charlie: "IKne3meq5aSn9XLyUdCD",
-  george: "JBFqnCBsd6RMkjVDRZzb",
-  emily: "LcfcDJNUP1GQjkzn1xUU",
-  elli: "MF3mGyEYCl7XYWbV9V6O",
-  callum: "N2lVS1w4EtoT3dr4eOWO",
-  patrick: "ODq5zmih8GrVes37Dizd",
-  harry: "SOYHLrjzK2X1ezoPC6cr",
-  liam: "TX3LPaxmHKxFdv7VOQHJ",
-  dorothy: "ThT5KcBeYPX3keUQqHPh",
-  josh: "TxGEqnHWrfWFTfGW9XjX",
-  arnold: "VR6AewLTigWG4xSOukaG",
-  charlotte: "XB0fDUnXU5powFXDhCwa",
-  matilda: "XrExE9yKIg1WjnnlVkGX",
-  matthew: "Yko7PKHZNXotIFUBG7I9",
-  james: "ZQe5CZNOzWyzPSCn5a3c",
-  joseph: "Zlb1dXrM653N07WRdFW3",
-  jeremy: "bVMeCyTHy58xNoL34h3p",
-  michael: "flq6f7yk4E4fJM5XTYuZ",
-  ethan: "g5CIjZEefAph4nQFvHAz",
-  gigi: "jBpfuIE2acCO8z3wKNLl",
-  freya: "jsCqWAovK2LkecY7zXl4",
-  brian: "nPczCjzI2devNBz1zQrb",
-  grace: "oWAxZDx7w5VEj9dCyTzz",
-  daniel: "onwK4e9ZLuTAKqWW03F9",
-  lily: "pFZP5JQG7iQjIQuC4Bku",
-  serena: "pMsXgVXv3BLzUgSXRplE",
-  adam: "pNInz6obpgDQGcFmaJgB",
-  nicole: "piTKgcLEGmPE4e6mEKli",
-  bill: "pqHfZKP75CvOlQylNhV4",
-  jessie: "t0jbNlBVZ17f02VDIeMI",
-  ryan: "wViXBPUzp2ZZixB1xQuM",
-  sam: "yoZ06aMxZJJ28mfd3POQ",
-  glinda: "z9fAnlkpzviPz146aGWa",
-  giovanni: "zcAOhNBS3c14rBihAFp1",
-  mimi: "zrHiDhphv9ZnVXBqCLjz",
-  sarah: "EXAVITQu4vr4xnSDxMaL",
-  alloy: "EXAVITQu4vr4xnSDxMaL"
-};
-var VOICE_ID_PATTERN = /^[A-Za-z0-9]{20}$/;
-function resolveVoiceId(voice) {
-  if (!voice) return voice;
-  if (VOICE_ID_PATTERN.test(voice)) return voice;
-  return ELEVENLABS_VOICE_ID_BY_NAME[voice.toLowerCase()] ?? voice;
-}
-var ElevenLabsModel = {
-  V3: "eleven_v3",
-  FLASH_V2_5: "eleven_flash_v2_5",
-  TURBO_V2_5: "eleven_turbo_v2_5",
-  MULTILINGUAL_V2: "eleven_multilingual_v2",
-  MONOLINGUAL_V1: "eleven_monolingual_v1"
-};
-var ElevenLabsOutputFormat = {
-  MP3_22050_32: "mp3_22050_32",
-  MP3_44100_32: "mp3_44100_32",
-  MP3_44100_64: "mp3_44100_64",
-  MP3_44100_96: "mp3_44100_96",
-  MP3_44100_128: "mp3_44100_128",
-  MP3_44100_192: "mp3_44100_192",
-  PCM_8000: "pcm_8000",
-  PCM_16000: "pcm_16000",
-  PCM_22050: "pcm_22050",
-  PCM_24000: "pcm_24000",
-  PCM_44100: "pcm_44100",
-  ULAW_8000: "ulaw_8000"
-};
-var ElevenLabsTTS = class _ElevenLabsTTS {
-  // Stable pricing/dashboard key — read by stream-handler / metrics via
-  // ``(agent.tts.constructor as any).providerKey``. Without this the cost
-  // calculator falls back to ``constructor.name`` ("ElevenLabsTTS") which
-  // does NOT match the pricing table key "elevenlabs", silently zeroing
-  // TTS cost for callers that construct the raw REST class directly
-  // (exposed at top level as ``ElevenLabsRestTTS``).
-  static providerKey = "elevenlabs";
-  apiKey;
-  voiceId;
-  modelId;
-  outputFormat;
-  voiceSettings;
-  languageCode;
-  chunkSize;
-  constructor(apiKey, voiceIdOrOptions = "21m00Tcm4TlvDq8ikWAM", modelId = ElevenLabsModel.FLASH_V2_5, outputFormat = ElevenLabsOutputFormat.PCM_16000) {
-    this.apiKey = apiKey;
-    if (typeof voiceIdOrOptions === "object") {
-      const o = voiceIdOrOptions;
-      this.voiceId = resolveVoiceId(o.voiceId ?? "21m00Tcm4TlvDq8ikWAM");
-      this.modelId = o.modelId ?? ElevenLabsModel.FLASH_V2_5;
-      this.outputFormat = o.outputFormat ?? ElevenLabsOutputFormat.PCM_16000;
-      this.voiceSettings = o.voiceSettings;
-      this.languageCode = o.languageCode;
-      this.chunkSize = o.chunkSize ?? 4096;
-    } else {
-      this.voiceId = resolveVoiceId(voiceIdOrOptions);
-      this.modelId = modelId;
-      this.outputFormat = outputFormat;
-      this.voiceSettings = void 0;
-      this.languageCode = void 0;
-      this.chunkSize = 4096;
-    }
-  }
-  /**
-   * Construct an instance pre-configured for Twilio Media Streams.
-   *
-   * Sets `outputFormat='ulaw_8000'` so ElevenLabs emits μ-law @ 8 kHz
-   * directly — the exact wire format Twilio's media stream uses — letting
-   * the SDK skip the 16 kHz→8 kHz resample and PCM→μ-law conversion in
-   * `TwilioAudioSender`. Saves ~30–80 ms first-byte and per-frame CPU,
-   * and removes a potential aliasing source.
-   *
-   * `voiceSettings` defaults to a low-bandwidth-friendly profile
-   * (speaker boost off, modest stability) which sounds cleaner at 8 kHz
-   * μ-law than the studio default. Pass an explicit object to override.
-   */
-  static forTwilio(apiKey, options = {}) {
-    const voiceSettings = options.voiceSettings ?? {
-      // Speaker boost adds high-frequency emphasis that aliases ugly over an
-      // 8 kHz μ-law line. Slightly higher stability tames the excursions
-      // that compander quantization noise can amplify.
-      stability: 0.6,
-      similarity_boost: 0.75,
-      use_speaker_boost: false
-    };
-    return new _ElevenLabsTTS(apiKey, {
-      ...options,
-      voiceSettings,
-      outputFormat: ElevenLabsOutputFormat.ULAW_8000
-    });
-  }
-  /**
-   * Construct an instance pre-configured for Telnyx bidirectional media.
-   *
-   * Telnyx's default media-streaming codec is L16 PCM @ 16 kHz, which
-   * matches our default Telnyx handler. We pick `pcm_16000` so the audio
-   * flows end-to-end with zero resampling or transcoding.
-   *
-   * Trade-off: if your Telnyx profile is pinned to PCMU/8000 (μ-law),
-   * construct `ElevenLabsTTS` directly with `outputFormat: 'ulaw_8000'`
-   * — Telnyx supports that natively too.
-   */
-  static forTelnyx(apiKey, options = {}) {
-    return new _ElevenLabsTTS(apiKey, {
-      ...options,
-      outputFormat: ElevenLabsOutputFormat.PCM_16000
-    });
-  }
-  /**
-   * Synthesise text to speech and return the full audio as a single Buffer.
-   *
-   * For large chunks (or when latency matters) call `synthesizeStream` instead.
-   */
-  async synthesize(text) {
-    const chunks = [];
-    for await (const chunk of this.synthesizeStream(text)) {
-      chunks.push(chunk);
-    }
-    return Buffer.concat(chunks);
-  }
-  /**
-   * Synthesise text and yield audio chunks as they arrive (streaming).
-   *
-   * The yielded buffers are raw PCM at 16 kHz (or whatever `outputFormat` is
-   * configured to). `chunkSize` controls the maximum yield size — 512 is a
-   * good choice for low-latency telephony.
-   */
-  async *synthesizeStream(text) {
-    const url = `${ELEVENLABS_BASE_URL}/text-to-speech/${encodeURIComponent(this.voiceId)}/stream?output_format=${encodeURIComponent(this.outputFormat)}`;
-    const body = {
-      text,
-      model_id: this.modelId
-    };
-    if (this.voiceSettings) body["voice_settings"] = this.voiceSettings;
-    if (this.languageCode) body["language_code"] = this.languageCode;
-    const response = await fetch(url, {
-      method: "POST",
-      headers: {
-        "xi-api-key": this.apiKey,
-        "Content-Type": "application/json"
-      },
-      body: JSON.stringify(body),
-      signal: AbortSignal.timeout(3e4)
-    });
-    if (!response.ok) {
-      const errBody = await response.text();
-      throw new Error(`ElevenLabs TTS error ${response.status}: ${errBody}`);
-    }
-    if (!response.body) {
-      throw new Error("ElevenLabs TTS: no response body");
-    }
-    const reader = response.body.getReader();
-    try {
-      while (true) {
-        const { done, value } = await reader.read();
-        if (done) break;
-        if (!value || value.length === 0) continue;
-        const buf = Buffer.from(value);
-        for (let offset = 0; offset < buf.length; offset += this.chunkSize) {
-          yield buf.subarray(offset, Math.min(offset + this.chunkSize, buf.length));
+    if (!ws) return;
+    this.ws = null;
+    const sendSafe = (payload) => {
+      if (ws.readyState === WebSocket5.OPEN) {
+        try {
+          ws.send(payload);
+        } catch {
         }
       }
-    } finally {
-      if (typeof reader.cancel === "function") await reader.cancel().catch(() => {
-      });
-      reader.releaseLock();
+    };
+    sendSafe(
+      JSON.stringify({ message: "EndOfStream", last_seq_no: this.lastSeqNo })
+    );
+    try {
+      ws.close();
+    } catch {
+    }
+  }
+};
+
+// src/stt/speechmatics.ts
+var STT7 = class extends SpeechmaticsSTT {
+  static providerKey = "speechmatics";
+  constructor(opts = {}) {
+    const key = opts.apiKey ?? process.env.SPEECHMATICS_API_KEY;
+    if (!key) {
+      throw new Error(
+        "Speechmatics STT requires an apiKey. Pass { apiKey: 'sm_...' } or set SPEECHMATICS_API_KEY in the environment."
+      );
     }
+    super(key, opts);
   }
 };
 
 // src/tts/elevenlabs.ts
+init_esm_shims();
 function resolveApiKey(apiKey) {
   const key = apiKey ?? process.env.ELEVENLABS_API_KEY;
   if (!key) {
@@ -4721,7 +5172,7 @@ var TTS = class _TTS extends ElevenLabsTTS {
     super(resolveApiKey(opts.apiKey), {
       voiceId: opts.voiceId ?? "EXAVITQu4vr4xnSDxMaL",
       modelId: opts.modelId ?? "eleven_flash_v2_5",
-      outputFormat: opts.outputFormat ?? "pcm_16000",
+      ...opts.outputFormat !== void 0 ? { outputFormat: opts.outputFormat } : {},
       languageCode: opts.languageCode,
       voiceSettings: opts.voiceSettings
     });
@@ -4792,6 +5243,20 @@ var ElevenLabsWebSocketTTS = class _ElevenLabsWebSocketTTS {
    * changes.
    */
   adoptedConnection = null;
+  /**
+   * Active WS for the in-flight ``synthesizeStream`` call, if any. Set
+   * when a stream starts, cleared in its ``finally`` block. The
+   * stream-handler calls ``cancelActiveStream()`` from ``cancelSpeaking``
+   * to unblock the generator's inner ``await Promise<frame>`` — without
+   * it, a barge-in on the firstMessage live path leaves the for-await
+   * stuck waiting for the next frame; ElevenLabs never sends
+   * ``isFinal=true`` after the consumer breaks, the 30 s frame timeout
+   * fires post-call, and meanwhile ``initPipeline`` never returns so
+   * the STT ``onTranscript`` callback never registers and subsequent
+   * user turns are silently dropped (root cause of the 2026-05-20
+   * "first message OK, then no response" symptom).
+   */
+  activeStreamWs = null;
   /**
    * The wire format requested over the ElevenLabs WS. Initially set from
    * the constructor; ``setTelephonyCarrier`` may auto-flip it to the
@@ -4840,6 +5305,32 @@ var ElevenLabsWebSocketTTS = class _ElevenLabsWebSocketTTS {
     if (!native) return;
     this._outputFormat = native;
   }
+  /**
+   * Force-close the WebSocket of any in-flight ``synthesizeStream`` call.
+   * Called by the stream-handler from ``cancelSpeaking`` (barge-in) so
+   * the generator's inner ``await Promise<frame>`` loop unblocks cleanly
+   * via the ``onClose`` handler — instead of waiting up to 30 s for the
+   * ``FRAME_TIMEOUT_MS`` watchdog to fire. No-op when no stream is in
+   * flight or when the WS is already closing.
+   *
+   * Without this, a barge-in during the firstMessage live path left the
+   * for-await stuck (ElevenLabs never sends ``isFinal=true`` after the
+   * consumer breaks), ``initPipeline`` never returned, the STT
+   * ``onTranscript`` callback never registered, and the entire remainder
+   * of the call was silent for the user. Surfaced during the 2026-05-20
+   * acceptance run.
+   */
+  cancelActiveStream() {
+    const ws = this.activeStreamWs;
+    if (!ws) return;
+    this.activeStreamWs = null;
+    try {
+      if (ws.readyState === WebSocket6.OPEN || ws.readyState === WebSocket6.CONNECTING) {
+        ws.close();
+      }
+    } catch {
+    }
+  }
   /** Pre-configured for Twilio Media Streams (`ulaw_8000`). */
   static forTwilio(opts) {
     return new _ElevenLabsWebSocketTTS({
@@ -4925,6 +5416,7 @@ var ElevenLabsWebSocketTTS = class _ElevenLabsWebSocketTTS {
         headers: { "xi-api-key": this.apiKey }
       });
     }
+    this.activeStreamWs = ws;
     const queue = [];
     let done = false;
     let pendingError = null;
@@ -5045,6 +5537,7 @@ var ElevenLabsWebSocketTTS = class _ElevenLabsWebSocketTTS {
       }
     } finally {
       if (connectTimer) clearTimeout(connectTimer);
+      if (this.activeStreamWs === ws) this.activeStreamWs = null;
       try {
         if (ws.readyState === WebSocket6.OPEN) {
           ws.send(JSON.stringify({ text: "" }));
@@ -5217,9 +5710,9 @@ function buildOpts(opts) {
   const out = {
     apiKey: resolveApiKey2(opts.apiKey),
     modelId: opts.modelId ?? "eleven_flash_v2_5",
-    outputFormat: opts.outputFormat ?? "pcm_16000",
     autoMode: opts.autoMode ?? true
   };
+  if (opts.outputFormat !== void 0) out.outputFormat = opts.outputFormat;
   if (opts.voiceId !== void 0) out.voiceId = opts.voiceId;
   if (opts.voiceSettings !== void 0) out.voiceSettings = opts.voiceSettings;
   if (opts.languageCode !== void 0) out.languageCode = opts.languageCode;
@@ -5396,268 +5889,77 @@ var OpenAITTS = class _OpenAITTS {
       if (lpf) {
         y = lpfAlpha * x + (1 - lpfAlpha) * y;
         let s = Math.round(y);
-        if (s > 32767) s = 32767;
-        else if (s < -32768) s = -32768;
-        samples.push(s);
-      } else {
-        samples.push(x);
-      }
-    }
-    if (lpf) ctx.lpfPrev = y;
-    const out = [];
-    let i = 0;
-    if (direct8k) {
-      while (i + 2 < samples.length) {
-        out.push(samples[i]);
-        i += 3;
-      }
-    } else {
-      while (i + 2 < samples.length) {
-        out.push(samples[i]);
-        out.push(Math.round((samples[i + 1] + samples[i + 2]) / 2));
-        i += 3;
-      }
-    }
-    ctx.leftover = samples.slice(i);
-    const buffer = Buffer.alloc(out.length * 2);
-    for (let j = 0; j < out.length; j++) {
-      buffer.writeInt16LE(out[j], j * 2);
-    }
-    return buffer;
-  }
-  /** @deprecated use {@link resampleStreaming} with persistent state. */
-  static resample24kTo16k(audio) {
-    const ctx = {
-      carryByte: null,
-      leftover: [],
-      lpfPrev: 0,
-      lpfEnabled: false,
-      targetSampleRate: 16e3
-    };
-    const out = _OpenAITTS.resampleStreaming(audio, ctx);
-    if (ctx.leftover.length === 0) return out;
-    const tail = Buffer.alloc(ctx.leftover.length * 2);
-    for (let i = 0; i < ctx.leftover.length; i++) {
-      tail.writeInt16LE(ctx.leftover[i], i * 2);
-    }
-    return Buffer.concat([out, tail]);
-  }
-};
-
-// src/tts/openai.ts
-var TTS3 = class extends OpenAITTS {
-  static providerKey = "openai_tts";
-  constructor(opts = {}) {
-    const key = opts.apiKey ?? process.env.OPENAI_API_KEY;
-    if (!key) {
-      throw new Error(
-        "OpenAI TTS requires an apiKey. Pass { apiKey: 'sk-...' } or set OPENAI_API_KEY in the environment."
-      );
-    }
-    super(
-      key,
-      opts.voice ?? "alloy",
-      opts.model ?? "gpt-4o-mini-tts",
-      opts.instructions ?? null,
-      opts.speed ?? null,
-      opts.antiAlias ?? false
-    );
-  }
-};
-
-// src/tts/cartesia.ts
-init_esm_shims();
-
-// src/providers/cartesia-tts.ts
-init_esm_shims();
-var CARTESIA_BASE_URL = "https://api.cartesia.ai";
-var CARTESIA_API_VERSION = "2025-04-16";
-var CARTESIA_DEFAULT_VOICE_ID = "f786b574-daa5-4673-aa0c-cbe3e8534c02";
-var CartesiaTTSModel = {
-  SONIC_3: "sonic-3",
-  SONIC_2: "sonic-2",
-  SONIC: "sonic"
-};
-var CartesiaTTSContainer = {
-  RAW: "raw",
-  WAV: "wav",
-  MP3: "mp3"
-};
-var CartesiaTTSEncoding = {
-  PCM_S16LE: "pcm_s16le",
-  PCM_F32LE: "pcm_f32le",
-  PCM_MULAW: "pcm_mulaw",
-  PCM_ALAW: "pcm_alaw"
-};
-var CartesiaTTSSampleRate = {
-  HZ_8000: 8e3,
-  HZ_16000: 16e3,
-  HZ_22050: 22050,
-  HZ_24000: 24e3,
-  HZ_44100: 44100
-};
-var CartesiaTTSVoiceMode = {
-  ID: "id",
-  EMBEDDING: "embedding"
-};
-var CartesiaTTS = class _CartesiaTTS {
-  /** Stable pricing/dashboard key — read by stream-handler/metrics. */
-  static providerKey = "cartesia_tts";
-  apiKey;
-  model;
-  voice;
-  language;
-  sampleRate;
-  speed;
-  emotion;
-  volume;
-  baseUrl;
-  apiVersion;
-  constructor(apiKey, opts = {}) {
-    this.apiKey = apiKey;
-    this.model = opts.model ?? CartesiaTTSModel.SONIC_3;
-    this.voice = opts.voice ?? CARTESIA_DEFAULT_VOICE_ID;
-    this.language = opts.language ?? "en";
-    this.sampleRate = opts.sampleRate ?? CartesiaTTSSampleRate.HZ_16000;
-    this.speed = opts.speed;
-    this.emotion = typeof opts.emotion === "string" ? [opts.emotion] : opts.emotion;
-    this.volume = opts.volume;
-    this.baseUrl = opts.baseUrl ?? CARTESIA_BASE_URL;
-    this.apiVersion = opts.apiVersion ?? CARTESIA_API_VERSION;
-  }
-  /**
-   * Construct an instance pre-configured for Twilio Media Streams.
-   *
-   * Sets `sampleRate=8000` so Cartesia emits PCM_S16LE @ 8 kHz directly.
-   * Twilio's media stream uses μ-law @ 8 kHz so the SDK still does the
-   * PCM → μ-law transcode client-side, but the 16 kHz → 8 kHz resample
-   * step is skipped. Saves ~10–30 ms first-byte plus per-frame CPU and
-   * removes a potential aliasing source.
-   */
-  static forTwilio(apiKey, options = {}) {
-    return new _CartesiaTTS(apiKey, {
-      ...options,
-      sampleRate: CartesiaTTSSampleRate.HZ_8000
-    });
-  }
-  /**
-   * Construct an instance pre-configured for Telnyx bidirectional media.
-   *
-   * Sets `sampleRate=16000` to match Telnyx's L16/16000 default codec —
-   * audio flows end-to-end with zero resampling or transcoding. Same as
-   * the bare-constructor default; exists for API symmetry with
-   * {@link CartesiaTTS.forTwilio}.
-   */
-  static forTelnyx(apiKey, options = {}) {
-    return new _CartesiaTTS(apiKey, {
-      ...options,
-      sampleRate: CartesiaTTSSampleRate.HZ_16000
-    });
-  }
-  /** Build the JSON payload for the Cartesia bytes endpoint. */
-  buildPayload(text) {
-    const payload = {
-      model_id: this.model,
-      voice: { mode: CartesiaTTSVoiceMode.ID, id: this.voice },
-      transcript: text,
-      output_format: {
-        container: CartesiaTTSContainer.RAW,
-        encoding: CartesiaTTSEncoding.PCM_S16LE,
-        sample_rate: this.sampleRate
-      },
-      language: this.language
-    };
-    const generationConfig = {};
-    if (this.speed !== void 0) generationConfig.speed = this.speed;
-    if (this.emotion && this.emotion.length > 0)
-      generationConfig.emotion = this.emotion[0];
-    if (this.volume !== void 0) generationConfig.volume = this.volume;
-    if (Object.keys(generationConfig).length > 0) {
-      payload.generation_config = generationConfig;
-    }
-    return payload;
-  }
-  /**
-   * Pre-call HTTP warmup for the Cartesia `/tts/bytes` endpoint.
-   *
-   * Issues a lightweight `GET <baseUrl>/voices` so DNS, TLS, and HTTP/2
-   * are already up by the time the first `synthesizeStream()` POST
-   * lands. Best-effort: 5 s timeout, all exceptions swallowed at
-   * debug level.
-   *
-   * Billing safety: `GET /voices` is a free metadata read on
-   * Cartesia's REST surface (per https://docs.cartesia.ai). It does
-   * not consume synthesis credits. The actual synthesis is billed
-   * only when `POST /tts/bytes` runs with a non-empty `transcript`.
-   *
-   * Note: Cartesia TTS uses the HTTP path (vs the WebSocket variant
-   * Cartesia also exposes) — connection warmup is therefore HTTP-GET
-   * based, not WebSocket pre-handshake. The latency win is smaller
-   * (~50-150 ms vs the ~200-500 ms of a WS prewarm) but still real.
-   */
-  async warmup() {
-    try {
-      await fetch(`${this.baseUrl}/voices`, {
-        method: "GET",
-        headers: {
-          "X-API-Key": this.apiKey,
-          "Cartesia-Version": this.apiVersion
-        },
-        signal: AbortSignal.timeout(5e3)
-      });
-    } catch (err) {
-      getLogger().debug(`Cartesia TTS warmup failed (best-effort): ${String(err)}`);
+        if (s > 32767) s = 32767;
+        else if (s < -32768) s = -32768;
+        samples.push(s);
+      } else {
+        samples.push(x);
+      }
     }
-  }
-  /** Synthesize text and return the concatenated audio buffer. */
-  async synthesize(text) {
-    const chunks = [];
-    for await (const chunk of this.synthesizeStream(text)) {
-      chunks.push(chunk);
+    if (lpf) ctx.lpfPrev = y;
+    const out = [];
+    let i = 0;
+    if (direct8k) {
+      while (i + 2 < samples.length) {
+        out.push(samples[i]);
+        i += 3;
+      }
+    } else {
+      while (i + 2 < samples.length) {
+        out.push(samples[i]);
+        out.push(Math.round((samples[i + 1] + samples[i + 2]) / 2));
+        i += 3;
+      }
     }
-    return Buffer.concat(chunks);
-  }
-  /**
-   * Synthesize text and yield raw PCM_S16LE chunks at the configured
-   * `sampleRate` as they arrive from Cartesia.
-   */
-  async *synthesizeStream(text) {
-    const response = await fetch(`${this.baseUrl}/tts/bytes`, {
-      method: "POST",
-      headers: {
-        "X-API-Key": this.apiKey,
-        "Cartesia-Version": this.apiVersion,
-        "Content-Type": "application/json"
-      },
-      body: JSON.stringify(this.buildPayload(text)),
-      signal: AbortSignal.timeout(3e4)
-    });
-    if (!response.ok) {
-      const body = await response.text();
-      throw new Error(`Cartesia TTS error ${response.status}: ${body}`);
+    ctx.leftover = samples.slice(i);
+    const buffer = Buffer.alloc(out.length * 2);
+    for (let j = 0; j < out.length; j++) {
+      buffer.writeInt16LE(out[j], j * 2);
     }
-    if (!response.body) {
-      throw new Error("Cartesia TTS: no response body");
+    return buffer;
+  }
+  /** @deprecated use {@link resampleStreaming} with persistent state. */
+  static resample24kTo16k(audio) {
+    const ctx = {
+      carryByte: null,
+      leftover: [],
+      lpfPrev: 0,
+      lpfEnabled: false,
+      targetSampleRate: 16e3
+    };
+    const out = _OpenAITTS.resampleStreaming(audio, ctx);
+    if (ctx.leftover.length === 0) return out;
+    const tail = Buffer.alloc(ctx.leftover.length * 2);
+    for (let i = 0; i < ctx.leftover.length; i++) {
+      tail.writeInt16LE(ctx.leftover[i], i * 2);
     }
-    const reader = response.body.getReader();
-    try {
-      while (true) {
-        const { done, value } = await reader.read();
-        if (done) break;
-        if (value && value.length > 0) {
-          yield Buffer.from(value);
-        }
-      }
-    } finally {
-      if (typeof reader.cancel === "function")
-        await reader.cancel().catch(() => {
-        });
-      reader.releaseLock();
+    return Buffer.concat([out, tail]);
+  }
+};
+
+// src/tts/openai.ts
+var TTS3 = class extends OpenAITTS {
+  static providerKey = "openai_tts";
+  constructor(opts = {}) {
+    const key = opts.apiKey ?? process.env.OPENAI_API_KEY;
+    if (!key) {
+      throw new Error(
+        "OpenAI TTS requires an apiKey. Pass { apiKey: 'sk-...' } or set OPENAI_API_KEY in the environment."
+      );
     }
+    super(
+      key,
+      opts.voice ?? "alloy",
+      opts.model ?? "gpt-4o-mini-tts",
+      opts.instructions ?? null,
+      opts.speed ?? null,
+      opts.antiAlias ?? false
+    );
   }
 };
 
 // src/tts/cartesia.ts
+init_esm_shims();
 function resolveApiKey3(apiKey) {
   const key = apiKey ?? process.env.CARTESIA_API_KEY;
   if (!key) {
@@ -5687,150 +5989,6 @@ var TTS4 = class _TTS extends CartesiaTTS {
 
 // src/tts/rime.ts
 init_esm_shims();
-
-// src/providers/rime-tts.ts
-init_esm_shims();
-var RIME_BASE_URL = "https://users.rime.ai/v1/rime-tts";
-var RimeModel = {
-  ARCANA: "arcana",
-  MIST: "mist",
-  MIST_V2: "mistv2"
-};
-var RimeAudioFormat = {
-  PCM: "audio/pcm",
-  MP3: "audio/mp3",
-  WAV: "audio/wav",
-  MULAW: "audio/mulaw"
-};
-var ARCANA_MODEL_TIMEOUT_MS = 60 * 4 * 1e3;
-var MIST_MODEL_TIMEOUT_MS = 30 * 1e3;
-function isMistModel(model) {
-  return model.includes(RimeModel.MIST);
-}
-function timeoutForModel(model) {
-  if (model === RimeModel.ARCANA) return ARCANA_MODEL_TIMEOUT_MS;
-  return MIST_MODEL_TIMEOUT_MS;
-}
-var RimeTTS = class {
-  /** Stable pricing/dashboard key — read by stream-handler/metrics. */
-  static providerKey = "rime";
-  apiKey;
-  model;
-  speaker;
-  lang;
-  sampleRate;
-  repetitionPenalty;
-  temperature;
-  topP;
-  maxTokens;
-  speedAlpha;
-  reduceLatency;
-  pauseBetweenBrackets;
-  phonemizeBetweenBrackets;
-  baseUrl;
-  totalTimeoutMs;
-  constructor(apiKey, opts = {}) {
-    this.apiKey = apiKey;
-    this.model = opts.model ?? RimeModel.ARCANA;
-    const defaultSpeaker = isMistModel(this.model) ? "cove" : "astra";
-    this.speaker = opts.speaker ?? defaultSpeaker;
-    this.lang = opts.lang ?? "eng";
-    this.sampleRate = opts.sampleRate ?? 16e3;
-    this.repetitionPenalty = opts.repetitionPenalty;
-    this.temperature = opts.temperature;
-    this.topP = opts.topP;
-    this.maxTokens = opts.maxTokens;
-    this.speedAlpha = opts.speedAlpha;
-    this.reduceLatency = opts.reduceLatency;
-    this.pauseBetweenBrackets = opts.pauseBetweenBrackets;
-    this.phonemizeBetweenBrackets = opts.phonemizeBetweenBrackets;
-    this.baseUrl = opts.baseUrl ?? RIME_BASE_URL;
-    this.totalTimeoutMs = timeoutForModel(this.model);
-  }
-  buildPayload(text) {
-    const payload = {
-      speaker: this.speaker,
-      text,
-      modelId: this.model
-    };
-    if (this.model === RimeModel.ARCANA) {
-      if (this.repetitionPenalty !== void 0)
-        payload.repetition_penalty = this.repetitionPenalty;
-      if (this.temperature !== void 0) payload.temperature = this.temperature;
-      if (this.topP !== void 0) payload.top_p = this.topP;
-      if (this.maxTokens !== void 0) payload.max_tokens = this.maxTokens;
-      payload.lang = this.lang;
-      payload.samplingRate = this.sampleRate;
-    } else if (isMistModel(this.model)) {
-      payload.lang = this.lang;
-      payload.samplingRate = this.sampleRate;
-      if (this.speedAlpha !== void 0) payload.speedAlpha = this.speedAlpha;
-      if (this.model === RimeModel.MIST_V2 && this.reduceLatency !== void 0) {
-        payload.reduceLatency = this.reduceLatency;
-      }
-      if (this.pauseBetweenBrackets !== void 0) {
-        payload.pauseBetweenBrackets = this.pauseBetweenBrackets;
-      }
-      if (this.phonemizeBetweenBrackets !== void 0) {
-        payload.phonemizeBetweenBrackets = this.phonemizeBetweenBrackets;
-      }
-    }
-    return payload;
-  }
-  /** Synthesize text and return the concatenated audio buffer. */
-  async synthesize(text) {
-    const chunks = [];
-    for await (const chunk of this.synthesizeStream(text)) {
-      chunks.push(chunk);
-    }
-    return Buffer.concat(chunks);
-  }
-  /**
-   * Synthesize text and yield raw PCM_S16LE chunks at the configured
-   * `sampleRate` as they stream in.
-   */
-  async *synthesizeStream(text) {
-    const response = await fetch(this.baseUrl, {
-      method: "POST",
-      headers: {
-        accept: RimeAudioFormat.PCM,
-        Authorization: `Bearer ${this.apiKey}`,
-        "content-type": "application/json"
-      },
-      body: JSON.stringify(this.buildPayload(text)),
-      signal: AbortSignal.timeout(this.totalTimeoutMs)
-    });
-    if (!response.ok) {
-      const body = await response.text();
-      throw new Error(`Rime TTS error ${response.status}: ${body}`);
-    }
-    const contentType = response.headers.get("content-type") ?? "";
-    if (!contentType.startsWith("audio")) {
-      const body = await response.text();
-      throw new Error(`Rime returned non-audio response: ${body.slice(0, 500)}`);
-    }
-    if (!response.body) {
-      throw new Error("Rime TTS: no response body");
-    }
-    const reader = response.body.getReader();
-    try {
-      while (true) {
-        const { done, value } = await reader.read();
-        if (done) break;
-        if (value && value.length > 0) {
-          yield Buffer.from(value);
-        }
-      }
-    } finally {
-      if (typeof reader.cancel === "function")
-        await reader.cancel().catch(() => {
-        });
-      reader.releaseLock();
-    }
-  }
-};
-
-// src/tts/rime.ts
 var TTS5 = class extends RimeTTS {
   static providerKey = "rime";
   constructor(opts = {}) {
@@ -6469,12 +6627,6 @@ init_esm_shims();
 
 // src/providers/groq-llm.ts
 init_esm_shims();
-
-// src/version.ts
-init_esm_shims();
-var VERSION = "0.5.5";
-
-// src/providers/groq-llm.ts
 var GROQ_BASE_URL = "https://api.groq.com/openai/v1";
 var GroqModel = {
   LLAMA_3_3_70B_VERSATILE: "llama-3.3-70b-versatile",
@@ -8131,12 +8283,28 @@ var TwilioAdapter = class _TwilioAdapter {
     return { callSid: call.sid };
   }
   /**
-   * Build a minimal ``<Response><Connect><Stream url="..."/></Connect></Response>``
-   * TwiML document. Mirrors the Python adapter's ``generate_stream_twiml``.
+   * Build a ``<Response><Connect><Stream url="...">`` TwiML document.
+   *
+   * ``parameters`` is forwarded as ``<Parameter name="..." value="..."/>``
+   * children of ``<Stream>``. Twilio Media Streams strips query-string params
+   * from the ``<Stream url=...>`` before the WS handshake, so
+   * ``<Parameter>`` tags are the supported way to pre-populate
+   * ``start.customParameters`` on the WS ``start`` frame. Used by the
+   * inbound path to carry caller / callee through to the bridge.
+   *
+   * Mirrors the Python adapter's ``generate_stream_twiml``.
    */
-  static generateStreamTwiml(streamUrl) {
-    const escaped = streamUrl.replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;").replace(/"/g, "&quot;").replace(/'/g, "&apos;");
-    return `<?xml version="1.0" encoding="UTF-8"?><Response><Connect><Stream url="${escaped}"/></Connect></Response>`;
+  static generateStreamTwiml(streamUrl, parameters) {
+    const esc = (s) => s.replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;").replace(/"/g, "&quot;").replace(/'/g, "&apos;");
+    const escapedUrl = esc(streamUrl);
+    let paramTags = "";
+    if (parameters) {
+      for (const [name, value] of Object.entries(parameters)) {
+        if (value == null) continue;
+        paramTags += `<Parameter name="${esc(name)}" value="${esc(String(value))}"/>`;
+      }
+    }
+    return `<?xml version="1.0" encoding="UTF-8"?><Response><Connect><Stream url="${escapedUrl}">${paramTags}</Stream></Connect></Response>`;
   }
   /** Force-complete an in-progress call. */
   async endCall(callSid) {
@@ -8529,6 +8697,8 @@ export {
   CallMetricsAccumulator,
   STT4 as CartesiaSTT,
   TTS4 as CartesiaTTS,
+  CartesiaTTSModel,
+  CartesiaTTSVoiceMode,
   LLM4 as CerebrasLLM,
   ChatContext,
   CloudflareTunnel,
@@ -8536,10 +8706,13 @@ export {
   DEFAULT_PRICING,
   DTMF_EVENTS,
   DeepFilterNetFilter,
+  DeepgramModel,
   STT as DeepgramSTT,
   DefaultToolExecutor,
   ConvAI as ElevenLabsConvAI,
   ElevenLabsConvAIAdapter,
+  ElevenLabsModel,
+  ElevenLabsOutputFormat,
   ElevenLabsTTS as ElevenLabsRestTTS,
   TTS as ElevenLabsTTS,
   TTS2 as ElevenLabsWebSocketTTS,
@@ -8568,8 +8741,15 @@ export {
   Realtime2 as OpenAIRealtime2,
   OpenAIRealtime2Adapter,
   OpenAIRealtimeAdapter,
+  OpenAIRealtimeAudioFormat,
+  OpenAIRealtimeModel,
+  OpenAIRealtimeVADType,
   TTS3 as OpenAITTS,
   STT3 as OpenAITranscribeSTT,
+  OpenAITranscriptionModel,
+  OpenAIVoice,
+  PRICING_LAST_UPDATED,
+  PRICING_VERSION,
   PartialStreamError,
   Patter,
   PatterConnectionError,
@@ -8577,9 +8757,12 @@ export {
   PatterTool,
   PcmCarry,
   PipelineHookExecutor,
+  PricingUnit,
   ProvisionError,
   RateLimitError,
   RemoteMessageHandler,
+  RimeAudioFormat,
+  RimeModel,
   TTS5 as RimeTTS,
   SPAN_BARGEIN,
   SPAN_CALL,