@livekit/agents-plugin-sarvam 1.0.45

package/dist/stt.d.cts

@@ -0,0 +1,104 @@
+import { type APIConnectOptions, type AudioBuffer, stt } from '@livekit/agents';
+import type { STTLanguages, STTModels, STTModes, STTV2Languages, STTV3Languages } from './models.js';
+interface STTBaseOptions {
+    /** Sarvam API key. Defaults to $SARVAM_API_KEY */
+    apiKey?: string;
+    /**
+     * Whether to use native WebSocket streaming for `stream()`.
+     * Set to `false` to prefer non-streaming REST recognition (used by Agent via StreamAdapter + VAD).
+     * Default: `true`.
+     */
+    streaming?: boolean;
+    /** Increase VAD sensitivity (WS only). Maps to `high_vad_sensitivity` query param. */
+    highVadSensitivity?: boolean;
+    /** Enable flush signal events from server (WS only). Maps to `flush_signal` query param. */
+    flushSignal?: boolean;
+}
+/**
+ * Options specific to saarika:v2.5.
+ * saarika:v2.5 will be deprecated soon — prefer {@link STTV3Options} with `saaras:v3` for new integrations.
+ * All v2.5 language codes are also supported by v3.
+ * @see {@link https://docs.sarvam.ai/api-reference-docs/speech-to-text/transcribe | Sarvam STT API docs}
+ */
+export interface STTV2Options extends STTBaseOptions {
+    model: 'saarika:v2.5';
+    /** Language code (BCP-47). Default: 'en-IN'. Set to 'unknown' for auto-detection. */
+    languageCode?: STTV2Languages | string;
+    /** Return chunk-level timestamps in REST response */
+    withTimestamps?: boolean;
+}
+/**
+ * Options specific to saaras:v2.5 (dedicated translate endpoint).
+ * Uses the `/speech-to-text-translate` endpoint for Indic-to-English translation.
+ * Auto-detects the source language; does not accept language codes or timestamps.
+ * @see {@link https://docs.sarvam.ai/api-reference-docs/speech-to-text-translate/translate | Sarvam STT Translate docs}
+ */
+export interface STTTranslateOptions extends STTBaseOptions {
+    model: 'saaras:v2.5';
+    /** Conversation context to boost model accuracy */
+    prompt?: string;
+    /** Mode for translate WS. Default: 'translate'. */
+    mode?: STTModes | string;
+}
+/**
+ * Options specific to saaras:v3 (recommended).
+ * @see {@link https://docs.sarvam.ai/api-reference-docs/speech-to-text/transcribe | Sarvam STT API docs}
+ */
+export interface STTV3Options extends STTBaseOptions {
+    model?: 'saaras:v3';
+    /** Language code (BCP-47). Default: 'en-IN'. Set to 'unknown' for auto-detection. */
+    languageCode?: STTV3Languages | string;
+    /** Transcription mode (v3 only). Default: 'transcribe' */
+    mode?: STTModes | string;
+    /** Conversation context to boost model accuracy */
+    prompt?: string;
+    /** Return chunk-level timestamps in REST response */
+    withTimestamps?: boolean;
+}
+/** Combined options — discriminated by `model` field */
+export type STTOptions = STTV2Options | STTTranslateOptions | STTV3Options;
+interface ResolvedSTTOptions {
+    apiKey: string;
+    model: STTModels;
+    streaming: boolean;
+    languageCode?: STTLanguages | string;
+    mode?: STTModes | string;
+    prompt?: string;
+    withTimestamps?: boolean;
+    highVadSensitivity?: boolean;
+    flushSignal?: boolean;
+}
+export declare class STT extends stt.STT {
+    private opts;
+    label: string;
+    /**
+     * Create a new instance of Sarvam AI STT.
+     *
+     * @remarks
+     * `apiKey` must be set to your Sarvam API key, either using the argument or by setting the
+     * `SARVAM_API_KEY` environment variable.
+     *
+     * Supported models:
+     * - `saaras:v3` (default, recommended) — supports all 22 languages, modes, prompt, timestamps, and uses `/speech-to-text`.
+     * - `saaras:v2.5` — Indic-to-English translation via `/speech-to-text-translate`. Auto-detects source language. Supports prompt.
+     * - `saarika:v2.5` — will be deprecated soon. Supports timestamps. All its languages are available in `saaras:v3`.
+     *
+     * @see {@link https://docs.sarvam.ai/api-reference-docs/speech-to-text/transcribe | Sarvam STT API docs}
+     * @see {@link https://docs.sarvam.ai/api-reference-docs/speech-to-text-translate/translate | Sarvam STT Translate docs}
+     */
+    constructor(opts?: Partial<STTOptions>);
+    updateOptions(opts: Partial<STTOptions>): void;
+    _recognize(buffer: AudioBuffer, abortSignal?: AbortSignal): Promise<stt.SpeechEvent>;
+    stream(options?: {
+        connOptions?: APIConnectOptions;
+    }): SpeechStream;
+}
+export declare class SpeechStream extends stt.SpeechStream {
+    #private;
+    label: string;
+    constructor(sttInstance: STT, opts: ResolvedSTTOptions, connOptions?: APIConnectOptions);
+    updateOptions(opts: Partial<STTOptions>): void;
+    protected run(): Promise<void>;
+}
+export {};
+//# sourceMappingURL=stt.d.ts.map

package/dist/stt.d.ts

@@ -0,0 +1,104 @@
+import { type APIConnectOptions, type AudioBuffer, stt } from '@livekit/agents';
+import type { STTLanguages, STTModels, STTModes, STTV2Languages, STTV3Languages } from './models.js';
+interface STTBaseOptions {
+    /** Sarvam API key. Defaults to $SARVAM_API_KEY */
+    apiKey?: string;
+    /**
+     * Whether to use native WebSocket streaming for `stream()`.
+     * Set to `false` to prefer non-streaming REST recognition (used by Agent via StreamAdapter + VAD).
+     * Default: `true`.
+     */
+    streaming?: boolean;
+    /** Increase VAD sensitivity (WS only). Maps to `high_vad_sensitivity` query param. */
+    highVadSensitivity?: boolean;
+    /** Enable flush signal events from server (WS only). Maps to `flush_signal` query param. */
+    flushSignal?: boolean;
+}
+/**
+ * Options specific to saarika:v2.5.
+ * saarika:v2.5 will be deprecated soon — prefer {@link STTV3Options} with `saaras:v3` for new integrations.
+ * All v2.5 language codes are also supported by v3.
+ * @see {@link https://docs.sarvam.ai/api-reference-docs/speech-to-text/transcribe | Sarvam STT API docs}
+ */
+export interface STTV2Options extends STTBaseOptions {
+    model: 'saarika:v2.5';
+    /** Language code (BCP-47). Default: 'en-IN'. Set to 'unknown' for auto-detection. */
+    languageCode?: STTV2Languages | string;
+    /** Return chunk-level timestamps in REST response */
+    withTimestamps?: boolean;
+}
+/**
+ * Options specific to saaras:v2.5 (dedicated translate endpoint).
+ * Uses the `/speech-to-text-translate` endpoint for Indic-to-English translation.
+ * Auto-detects the source language; does not accept language codes or timestamps.
+ * @see {@link https://docs.sarvam.ai/api-reference-docs/speech-to-text-translate/translate | Sarvam STT Translate docs}
+ */
+export interface STTTranslateOptions extends STTBaseOptions {
+    model: 'saaras:v2.5';
+    /** Conversation context to boost model accuracy */
+    prompt?: string;
+    /** Mode for translate WS. Default: 'translate'. */
+    mode?: STTModes | string;
+}
+/**
+ * Options specific to saaras:v3 (recommended).
+ * @see {@link https://docs.sarvam.ai/api-reference-docs/speech-to-text/transcribe | Sarvam STT API docs}
+ */
+export interface STTV3Options extends STTBaseOptions {
+    model?: 'saaras:v3';
+    /** Language code (BCP-47). Default: 'en-IN'. Set to 'unknown' for auto-detection. */
+    languageCode?: STTV3Languages | string;
+    /** Transcription mode (v3 only). Default: 'transcribe' */
+    mode?: STTModes | string;
+    /** Conversation context to boost model accuracy */
+    prompt?: string;
+    /** Return chunk-level timestamps in REST response */
+    withTimestamps?: boolean;
+}
+/** Combined options — discriminated by `model` field */
+export type STTOptions = STTV2Options | STTTranslateOptions | STTV3Options;
+interface ResolvedSTTOptions {
+    apiKey: string;
+    model: STTModels;
+    streaming: boolean;
+    languageCode?: STTLanguages | string;
+    mode?: STTModes | string;
+    prompt?: string;
+    withTimestamps?: boolean;
+    highVadSensitivity?: boolean;
+    flushSignal?: boolean;
+}
+export declare class STT extends stt.STT {
+    private opts;
+    label: string;
+    /**
+     * Create a new instance of Sarvam AI STT.
+     *
+     * @remarks
+     * `apiKey` must be set to your Sarvam API key, either using the argument or by setting the
+     * `SARVAM_API_KEY` environment variable.
+     *
+     * Supported models:
+     * - `saaras:v3` (default, recommended) — supports all 22 languages, modes, prompt, timestamps, and uses `/speech-to-text`.
+     * - `saaras:v2.5` — Indic-to-English translation via `/speech-to-text-translate`. Auto-detects source language. Supports prompt.
+     * - `saarika:v2.5` — will be deprecated soon. Supports timestamps. All its languages are available in `saaras:v3`.
+     *
+     * @see {@link https://docs.sarvam.ai/api-reference-docs/speech-to-text/transcribe | Sarvam STT API docs}
+     * @see {@link https://docs.sarvam.ai/api-reference-docs/speech-to-text-translate/translate | Sarvam STT Translate docs}
+     */
+    constructor(opts?: Partial<STTOptions>);
+    updateOptions(opts: Partial<STTOptions>): void;
+    _recognize(buffer: AudioBuffer, abortSignal?: AbortSignal): Promise<stt.SpeechEvent>;
+    stream(options?: {
+        connOptions?: APIConnectOptions;
+    }): SpeechStream;
+}
+export declare class SpeechStream extends stt.SpeechStream {
+    #private;
+    label: string;
+    constructor(sttInstance: STT, opts: ResolvedSTTOptions, connOptions?: APIConnectOptions);
+    updateOptions(opts: Partial<STTOptions>): void;
+    protected run(): Promise<void>;
+}
+export {};
+//# sourceMappingURL=stt.d.ts.map

package/dist/stt.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"stt.d.ts","sourceRoot":"","sources":["../src/stt.ts"],"names":[],"mappings":"AAGA,OAAO,EACL,KAAK,iBAAiB,EACtB,KAAK,WAAW,EAOhB,GAAG,EAEJ,MAAM,iBAAiB,CAAC;AAGzB,OAAO,KAAK,EACV,YAAY,EACZ,SAAS,EACT,QAAQ,EACR,cAAc,EACd,cAAc,EACf,MAAM,aAAa,CAAC;AAkBrB,UAAU,cAAc;IACtB,kDAAkD;IAClD,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB;;;;OAIG;IACH,SAAS,CAAC,EAAE,OAAO,CAAC;IACpB,sFAAsF;IACtF,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B,4FAA4F;IAC5F,WAAW,CAAC,EAAE,OAAO,CAAC;CACvB;AAED;;;;;GAKG;AACH,MAAM,WAAW,YAAa,SAAQ,cAAc;IAClD,KAAK,EAAE,cAAc,CAAC;IACtB,qFAAqF;IACrF,YAAY,CAAC,EAAE,cAAc,GAAG,MAAM,CAAC;IACvC,qDAAqD;IACrD,cAAc,CAAC,EAAE,OAAO,CAAC;CAC1B;AAED;;;;;GAKG;AACH,MAAM,WAAW,mBAAoB,SAAQ,cAAc;IACzD,KAAK,EAAE,aAAa,CAAC;IACrB,mDAAmD;IACnD,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,mDAAmD;IACnD,IAAI,CAAC,EAAE,QAAQ,GAAG,MAAM,CAAC;CAC1B;AAED;;;GAGG;AACH,MAAM,WAAW,YAAa,SAAQ,cAAc;IAClD,KAAK,CAAC,EAAE,WAAW,CAAC;IACpB,qFAAqF;IACrF,YAAY,CAAC,EAAE,cAAc,GAAG,MAAM,CAAC;IACvC,0DAA0D;IAC1D,IAAI,CAAC,EAAE,QAAQ,GAAG,MAAM,CAAC;IACzB,mDAAmD;IACnD,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,qDAAqD;IACrD,cAAc,CAAC,EAAE,OAAO,CAAC;CAC1B;AAED,wDAAwD;AACxD,MAAM,MAAM,UAAU,GAAG,YAAY,GAAG,mBAAmB,GAAG,YAAY,CAAC;AAM3E,UAAU,kBAAkB;IAC1B,MAAM,EAAE,MAAM,CAAC;IACf,KAAK,EAAE,SAAS,CAAC;IACjB,SAAS,EAAE,OAAO,CAAC;IAEnB,YAAY,CAAC,EAAE,YAAY,GAAG,MAAM,CAAC;IAErC,IAAI,CAAC,EAAE,QAAQ,GAAG,MAAM,CAAC;IAEzB,MAAM,CAAC,EAAE,MAAM,CAAC;IAEhB,cAAc,CAAC,EAAE,OAAO,CAAC;IAEzB,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B,WAAW,CAAC,EAAE,OAAO,CAAC;CACvB;AAgOD,qBAAa,GAAI,SAAQ,GAAG,CAAC,GAAG;IAC9B,OAAO,CAAC,IAAI,CAAqB;IACjC,KAAK,SAAgB;IAErB;;;;;;;;;;;;;;OAcG;gBACS,IAAI,GAAE,OAAO,CAAC,UAAU,CAAM;IAU1C,aAAa,CAAC,IAAI,EAAE,OAAO,CAAC,UAAU,CAAC;IAoBjC,UAAU,CAAC,MAAM,EAAE,WAAW,EAAE,WAAW,CAAC,EAAE,WAAW,GAAG,OAAO,CAAC,GAAG,CAAC,WAAW,CAAC;IA+C1F,MAAM,CAAC,OAAO,CAAC,EAAE;QAAE,WAAW,CAAC,EAAE,iBAAiB,CAAA;KAAE,GAAG,YAAY;CAQpE;AAMD,qBAAa,YAAa,SAAQ,GAAG,CAAC,YAAY;;IAOhD,KAAK,SAAyB;gBAElB,WAAW,EAAE,GAAG,EAAE,IAAI,EAAE,kBAAkB,EAAE,WAAW,CAAC,EAAE,iBAAiB;IAOvF,aAAa,CAAC,IAAI,EAAE,OAAO,CAAC,UAAU,CAAC;cAoBvB,GAAG;CAqRpB"}

package/dist/stt.js

@@ -0,0 +1,483 @@
+import {
+  AudioByteStream,
+  AudioEnergyFilter,
+  Future,
+  Task,
+  log,
+  mergeFrames,
+  stt,
+  waitForAbort
+} from "@livekit/agents";
+import { WebSocket } from "ws";
+const SARVAM_STT_REST_URL = "https://api.sarvam.ai/speech-to-text";
+const SARVAM_STT_TRANSLATE_REST_URL = "https://api.sarvam.ai/speech-to-text-translate";
+const SARVAM_STT_WS_URL = "wss://api.sarvam.ai/speech-to-text/ws";
+const SARVAM_STT_TRANSLATE_WS_URL = "wss://api.sarvam.ai/speech-to-text-translate/ws";
+const SAMPLE_RATE = 16e3;
+const NUM_CHANNELS = 1;
+const SAARIKA_DEFAULTS = {
+  languageCode: "en-IN"
+};
+const SAARAS_V3_DEFAULTS = {
+  languageCode: "en-IN",
+  mode: "transcribe"
+};
+const SAARAS_TRANSLATE_DEFAULTS = {
+  mode: "translate"
+};
+const STTV2_LANGUAGE_SET = /* @__PURE__ */ new Set([
+  "unknown",
+  "hi-IN",
+  "bn-IN",
+  "kn-IN",
+  "ml-IN",
+  "mr-IN",
+  "od-IN",
+  "pa-IN",
+  "ta-IN",
+  "te-IN",
+  "en-IN",
+  "gu-IN"
+]);
+function resolveOptions(opts) {
+  const apiKey = opts.apiKey ?? process.env.SARVAM_API_KEY;
+  if (!apiKey) {
+    throw new Error("Sarvam API key is required, whether as an argument or as $SARVAM_API_KEY");
+  }
+  const model = opts.model ?? "saaras:v3";
+  const base = {
+    apiKey,
+    model,
+    streaming: opts.streaming ?? true,
+    highVadSensitivity: opts.highVadSensitivity,
+    flushSignal: opts.flushSignal
+  };
+  if (model === "saaras:v2.5") {
+    const translateOpts = opts;
+    base.prompt = translateOpts.prompt;
+    base.mode = translateOpts.mode ?? SAARAS_TRANSLATE_DEFAULTS.mode;
+  } else if (model === "saaras:v3") {
+    const v3Opts = opts;
+    base.languageCode = v3Opts.languageCode ?? SAARAS_V3_DEFAULTS.languageCode;
+    base.mode = v3Opts.mode ?? SAARAS_V3_DEFAULTS.mode;
+    base.prompt = v3Opts.prompt;
+    base.withTimestamps = v3Opts.withTimestamps;
+  } else {
+    let languageCode = opts.languageCode ?? SAARIKA_DEFAULTS.languageCode;
+    if (!STTV2_LANGUAGE_SET.has(languageCode)) {
+      languageCode = SAARIKA_DEFAULTS.languageCode;
+    }
+    base.languageCode = languageCode;
+    base.withTimestamps = opts.withTimestamps;
+  }
+  return base;
+}
+function getRestUrl(model) {
+  return model === "saaras:v2.5" ? SARVAM_STT_TRANSLATE_REST_URL : SARVAM_STT_REST_URL;
+}
+function getWsUrl(model) {
+  return model === "saaras:v2.5" ? SARVAM_STT_TRANSLATE_WS_URL : SARVAM_STT_WS_URL;
+}
+function buildWsUrl(opts) {
+  const base = getWsUrl(opts.model);
+  const params = new URLSearchParams();
+  params.set("model", opts.model);
+  params.set("vad_signals", "true");
+  params.set("sample_rate", String(SAMPLE_RATE));
+  params.set("input_audio_codec", "pcm_s16le");
+  if (opts.model !== "saaras:v2.5" && opts.languageCode != null) {
+    params.set("language-code", opts.languageCode);
+  }
+  if (opts.mode != null) {
+    params.set("mode", opts.mode);
+  }
+  if (opts.highVadSensitivity != null) {
+    params.set("high_vad_sensitivity", String(opts.highVadSensitivity));
+  }
+  if (opts.flushSignal != null) {
+    params.set("flush_signal", String(opts.flushSignal));
+  }
+  return `${base}?${params.toString()}`;
+}
+function buildFormData(wavBlob, opts) {
+  const formData = new FormData();
+  formData.append("file", wavBlob, "audio.wav");
+  formData.append("model", opts.model);
+  if (opts.model !== "saaras:v2.5" && opts.languageCode != null) {
+    formData.append("language_code", opts.languageCode);
+  }
+  if (opts.model === "saaras:v3" && opts.mode != null) {
+    formData.append("mode", opts.mode);
+  }
+  if ((opts.model === "saaras:v2.5" || opts.model === "saaras:v3") && opts.prompt != null) {
+    formData.append("prompt", opts.prompt);
+  }
+  if (opts.model !== "saaras:v2.5" && opts.withTimestamps) {
+    formData.append("with_timestamps", "true");
+  }
+  return formData;
+}
+function createWav(frame) {
+  const bitsPerSample = 16;
+  const byteRate = frame.sampleRate * frame.channels * bitsPerSample / 8;
+  const blockAlign = frame.channels * bitsPerSample / 8;
+  const header = Buffer.alloc(44);
+  header.write("RIFF", 0);
+  header.writeUInt32LE(36 + frame.data.byteLength, 4);
+  header.write("WAVE", 8);
+  header.write("fmt ", 12);
+  header.writeUInt32LE(16, 16);
+  header.writeUInt16LE(1, 20);
+  header.writeUInt16LE(frame.channels, 22);
+  header.writeUInt32LE(frame.sampleRate, 24);
+  header.writeUInt32LE(byteRate, 28);
+  header.writeUInt16LE(blockAlign, 32);
+  header.writeUInt16LE(bitsPerSample, 34);
+  header.write("data", 36);
+  header.writeUInt32LE(frame.data.byteLength, 40);
+  const pcm = Buffer.from(frame.data.buffer, frame.data.byteOffset, frame.data.byteLength);
+  return Buffer.concat([header, pcm]);
+}
+class STT extends stt.STT {
+  opts;
+  label = "sarvam.STT";
+  /**
+   * Create a new instance of Sarvam AI STT.
+   *
+   * @remarks
+   * `apiKey` must be set to your Sarvam API key, either using the argument or by setting the
+   * `SARVAM_API_KEY` environment variable.
+   *
+   * Supported models:
+   * - `saaras:v3` (default, recommended) — supports all 22 languages, modes, prompt, timestamps, and uses `/speech-to-text`.
+   * - `saaras:v2.5` — Indic-to-English translation via `/speech-to-text-translate`. Auto-detects source language. Supports prompt.
+   * - `saarika:v2.5` — will be deprecated soon. Supports timestamps. All its languages are available in `saaras:v3`.
+   *
+   * @see {@link https://docs.sarvam.ai/api-reference-docs/speech-to-text/transcribe | Sarvam STT API docs}
+   * @see {@link https://docs.sarvam.ai/api-reference-docs/speech-to-text-translate/translate | Sarvam STT Translate docs}
+   */
+  constructor(opts = {}) {
+    const resolved = resolveOptions(opts);
+    super({
+      streaming: resolved.streaming,
+      interimResults: false,
+      alignedTranscript: false
+    });
+    this.opts = resolved;
+  }
+  updateOptions(opts) {
+    const modelChanging = opts.model != null && opts.model !== this.opts.model;
+    const base = modelChanging ? {
+      apiKey: this.opts.apiKey,
+      streaming: this.opts.streaming,
+      ...this.opts.highVadSensitivity != null ? { highVadSensitivity: this.opts.highVadSensitivity } : {},
+      ...this.opts.flushSignal != null ? { flushSignal: this.opts.flushSignal } : {},
+      ...this.opts.languageCode != null && opts.model !== "saaras:v2.5" ? { languageCode: this.opts.languageCode } : {}
+    } : { ...this.opts };
+    this.opts = resolveOptions({ ...base, ...opts });
+  }
+  async _recognize(buffer, abortSignal) {
+    const frame = mergeFrames(buffer);
+    const wavBuffer = createWav(frame);
+    const wavBlob = new Blob([new Uint8Array(wavBuffer)], { type: "audio/wav" });
+    const formData = buildFormData(wavBlob, this.opts);
+    const response = await fetch(getRestUrl(this.opts.model), {
+      method: "POST",
+      headers: {
+        "api-subscription-key": this.opts.apiKey
+      },
+      body: formData,
+      signal: abortSignal ?? null
+    });
+    if (!response.ok) {
+      const errorBody = await response.text();
+      throw new Error(`Sarvam STT API error ${response.status}: ${errorBody}`);
+    }
+    const data = await response.json();
+    let startTime = 0;
+    let endTime = 0;
+    if (data.timestamps) {
+      const starts = data.timestamps.start_time_seconds;
+      const ends = data.timestamps.end_time_seconds;
+      if (starts.length > 0) startTime = starts[0] ?? 0;
+      if (ends.length > 0) endTime = ends[ends.length - 1] ?? 0;
+    }
+    return {
+      type: stt.SpeechEventType.FINAL_TRANSCRIPT,
+      requestId: data.request_id ?? void 0,
+      alternatives: [
+        {
+          text: data.transcript || "",
+          language: data.language_code ?? this.opts.languageCode ?? "unknown",
+          startTime,
+          endTime,
+          confidence: data.language_probability ?? 0
+        }
+      ]
+    };
+  }
+  stream(options) {
+    if (!this.capabilities.streaming) {
+      throw new Error(
+        "Sarvam STT streaming is disabled (`streaming: false`). Use recognize() for REST or wrap with stt.StreamAdapter + VAD for streaming behavior."
+      );
+    }
+    return new SpeechStream(this, this.opts, options == null ? void 0 : options.connOptions);
+  }
+}
+class SpeechStream extends stt.SpeechStream {
+  #opts;
+  #audioEnergyFilter;
+  #logger = log();
+  #speaking = false;
+  #resetWS = new Future();
+  #requestId = "";
+  label = "sarvam.SpeechStream";
+  constructor(sttInstance, opts, connOptions) {
+    super(sttInstance, SAMPLE_RATE, connOptions);
+    this.#opts = opts;
+    this.closed = false;
+    this.#audioEnergyFilter = new AudioEnergyFilter();
+  }
+  updateOptions(opts) {
+    const modelChanging = opts.model != null && opts.model !== this.#opts.model;
+    const base = modelChanging ? {
+      apiKey: this.#opts.apiKey,
+      ...this.#opts.highVadSensitivity != null ? { highVadSensitivity: this.#opts.highVadSensitivity } : {},
+      ...this.#opts.flushSignal != null ? { flushSignal: this.#opts.flushSignal } : {},
+      ...this.#opts.languageCode != null && opts.model !== "saaras:v2.5" ? { languageCode: this.#opts.languageCode } : {}
+    } : { ...this.#opts };
+    this.#opts = resolveOptions({ ...base, ...opts });
+    this.#resetWS.resolve();
+  }
+  async run() {
+    const maxRetry = 32;
+    let retries = 0;
+    while (!this.input.closed && !this.closed) {
+      const wsUrl = buildWsUrl(this.#opts);
+      this.#logger.info(`Sarvam STT connecting to: ${wsUrl}`);
+      const ws = new WebSocket(wsUrl, {
+        headers: { "api-subscription-key": this.#opts.apiKey }
+      });
+      let sessionStart = 0;
+      try {
+        await new Promise((resolve, reject) => {
+          ws.once("open", () => resolve());
+          ws.once("error", (err) => reject(err));
+          ws.once(
+            "close",
+            (code) => reject(new Error(`WebSocket closed with code ${code}`))
+          );
+        });
+        sessionStart = Date.now();
+        await this.#runWS(ws);
+        retries = 0;
+      } catch (e) {
+        ws.removeAllListeners();
+        ws.close();
+        if (!this.closed && !this.input.closed) {
+          if (sessionStart > 0 && Date.now() - sessionStart > 5e3) {
+            retries = 0;
+          }
+          if (retries >= maxRetry) {
+            throw new Error(`Failed to connect to Sarvam STT after ${retries} attempts: ${e}`);
+          }
+          const delay = Math.min(retries * 5, 10);
+          retries++;
+          this.#logger.warn(
+            `Failed to connect to Sarvam STT, retrying in ${delay}s: ${e} (${retries}/${maxRetry})`
+          );
+          await new Promise((resolve) => setTimeout(resolve, delay * 1e3));
+        } else {
+          this.#logger.warn(
+            `Sarvam STT disconnected, connection is closed: ${e} (inputClosed: ${this.input.closed}, isClosed: ${this.closed})`
+          );
+        }
+      }
+    }
+    this.closed = true;
+  }
+  async #runWS(ws) {
+    this.#resetWS = new Future();
+    this.#speaking = false;
+    let closing = false;
+    const sessionController = new AbortController();
+    if (this.#opts.model === "saaras:v2.5" && this.#opts.prompt != null) {
+      ws.send(JSON.stringify({ type: "config", prompt: this.#opts.prompt }));
+    }
+    const wsMonitor = Task.from(async (controller) => {
+      const closed = new Promise((_, reject) => {
+        ws.once("close", (code, reason) => {
+          if (!closing) {
+            this.#logger.error(`WebSocket closed with code ${code}: ${reason}`);
+            reject(new Error("WebSocket closed"));
+          }
+        });
+      });
+      await Promise.race([closed, waitForAbort(controller.signal)]);
+    });
+    const sendTask = async () => {
+      const samples50Ms = Math.floor(SAMPLE_RATE / 20);
+      const stream = new AudioByteStream(SAMPLE_RATE, NUM_CHANNELS, samples50Ms);
+      const abortPromise = waitForAbort(this.abortSignal);
+      const sessionAbort = waitForAbort(sessionController.signal);
+      try {
+        while (!this.closed) {
+          const result = await Promise.race([this.input.next(), abortPromise, sessionAbort]);
+          if (result === void 0) return;
+          if (result.done) break;
+          const data = result.value;
+          let frames;
+          if (data === SpeechStream.FLUSH_SENTINEL) {
+            frames = stream.flush();
+          } else if (data.sampleRate !== SAMPLE_RATE || data.channels !== NUM_CHANNELS) {
+            throw new Error(
+              `Expected ${SAMPLE_RATE}Hz/${NUM_CHANNELS}ch, got ${data.sampleRate}Hz/${data.channels}ch`
+            );
+          } else {
+            frames = stream.write(
+              data.data.buffer.slice(
+                data.data.byteOffset,
+                data.data.byteOffset + data.data.byteLength
+              )
+            );
+          }
+          for (const frame of frames) {
+            if (this.#audioEnergyFilter.pushFrame(frame)) {
+              const pcmBuffer = Buffer.from(
+                frame.data.buffer,
+                frame.data.byteOffset,
+                frame.data.byteLength
+              );
+              const base64Audio = pcmBuffer.toString("base64");
+              ws.send(
+                JSON.stringify({
+                  audio: {
+                    data: base64Audio,
+                    encoding: "audio/wav",
+                    sample_rate: SAMPLE_RATE
+                  }
+                })
+              );
+            }
+          }
+          if (data === SpeechStream.FLUSH_SENTINEL) {
+            ws.send(JSON.stringify({ type: "flush" }));
+          }
+        }
+      } finally {
+        closing = true;
+        try {
+          ws.send(
+            JSON.stringify({
+              type: "end_of_stream",
+              audio: { data: "", encoding: "audio/wav", sample_rate: SAMPLE_RATE }
+            })
+          );
+        } catch {
+        }
+        wsMonitor.cancel();
+      }
+    };
+    const listenTask = Task.from(async (controller) => {
+      const putMessage = (event) => {
+        if (!this.queue.closed) {
+          try {
+            this.queue.put(event);
+          } catch {
+          }
+        }
+      };
+      const listenMessage = new Promise((resolve, reject) => {
+        ws.once("close", () => resolve());
+        ws.on("message", (msg) => {
+          var _a;
+          try {
+            const raw = msg.toString();
+            this.#logger.debug(`Sarvam STT raw WS message: ${raw.substring(0, 500)}`);
+            const json = JSON.parse(raw);
+            const msgType = json["type"] ?? "";
+            if (msgType === "events") {
+              const eventData = json["data"] ?? {};
+              const signalType = eventData.signal_type;
+              if (signalType === "START_SPEECH") {
+                if (!this.#speaking) {
+                  this.#speaking = true;
+                  putMessage({ type: stt.SpeechEventType.START_OF_SPEECH });
+                }
+              } else if (signalType === "END_SPEECH") {
+                if (this.#speaking) {
+                  this.#speaking = false;
+                  putMessage({ type: stt.SpeechEventType.END_OF_SPEECH });
+                }
+              }
+            } else if (msgType === "data") {
+              const td = json["data"] ?? {};
+              const transcript = td.transcript ?? "";
+              const language = td.language_code ?? this.#opts.languageCode ?? "unknown";
+              const requestId = td.request_id ?? "";
+              const confidence = td.language_probability ?? 0;
+              this.#requestId = requestId;
+              if (td.metrics) {
+                this.#logger.debug(
+                  `Sarvam STT metrics: audio_duration=${td.metrics.audio_duration}s, latency=${td.metrics.processing_latency}s`
+                );
+              }
+              if (transcript) {
+                if (!this.#speaking) {
+                  this.#speaking = true;
+                  putMessage({ type: stt.SpeechEventType.START_OF_SPEECH });
+                }
+                putMessage({
+                  type: stt.SpeechEventType.FINAL_TRANSCRIPT,
+                  requestId,
+                  alternatives: [
+                    {
+                      text: transcript,
+                      language,
+                      startTime: 0,
+                      endTime: ((_a = td.metrics) == null ? void 0 : _a.audio_duration) ?? 0,
+                      confidence
+                    }
+                  ]
+                });
+              }
+            } else if (msgType === "error") {
+              const nested = json["data"];
+              const errorInfo = (nested == null ? void 0 : nested.message) ?? (nested == null ? void 0 : nested.error) ?? json["error"] ?? json["message"] ?? "Unknown error";
+              const errorCode = (nested == null ? void 0 : nested.code) ?? json["code"] ?? "";
+              this.#logger.error(`Sarvam STT WebSocket error [${errorCode}]: ${errorInfo}`);
+              reject(new Error(`Sarvam STT API error [${errorCode}]: ${errorInfo}`));
+              return;
+            }
+            if (this.closed || closing) {
+              resolve();
+            }
+          } catch (err) {
+            this.#logger.error(`Error processing Sarvam STT message: ${msg}`);
+            reject(err);
+          }
+        });
+      });
+      await Promise.race([listenMessage, waitForAbort(controller.signal)]);
+    }, this.abortController);
+    try {
+      await Promise.race([
+        this.#resetWS.await,
+        Promise.all([sendTask(), listenTask.result, wsMonitor.result])
+      ]);
+    } finally {
+      closing = true;
+      sessionController.abort();
+      wsMonitor.cancel();
+      ws.close();
+      listenTask.result.catch(() => {
+      });
+    }
+  }
+}
+export {
+  STT,
+  SpeechStream
+};
+//# sourceMappingURL=stt.js.map