# # Copyright (c) 2024–2026, Daily # # SPDX-License-Identifier: BSD 2-Clause License # """Sarvam AI Speech-to-Text service implementation. This module provides a streaming Speech-to-Text service using Sarvam AI's WebSocket-based API. It supports real-time transcription with Voice Activity Detection (VAD) and can handle multiple audio formats for Indian language speech recognition. """ import base64 from dataclasses import dataclass, field from typing import Any, AsyncGenerator, Dict, Literal, Optional from loguru import logger from pydantic import BaseModel from pipecat.frames.frames import ( CancelFrame, EndFrame, ErrorFrame, Frame, StartFrame, TranscriptionFrame, UserStartedSpeakingFrame, UserStoppedSpeakingFrame, VADUserStartedSpeakingFrame, VADUserStoppedSpeakingFrame, ) from pipecat.processors.frame_processor import FrameDirection from pipecat.services.sarvam._sdk import sdk_headers from pipecat.services.settings import ( NOT_GIVEN, STTSettings, _NotGiven, _warn_deprecated_param, is_given, ) from pipecat.services.stt_latency import SARVAM_TTFS_P99 from pipecat.services.stt_service import STTService from pipecat.transcriptions.language import Language, resolve_language from pipecat.utils.time import time_now_iso8601 from pipecat.utils.tracing.service_decorators import traced_stt try: from sarvamai import AsyncSarvamAI from sarvamai.core.api_error import ApiError from sarvamai.core.events import EventType except ModuleNotFoundError as e: logger.error(f"Exception: {e}") logger.error("In order to use Sarvam, you need to `pip install pipecat-ai[sarvam]`.") raise Exception(f"Missing module: {e}") def language_to_sarvam_language(language: Language) -> str: """Convert a Language enum to Sarvam's language code format. Args: language: The Language enum value to convert. Returns: The Sarvam language code string. """ # Mapping of pipecat Language enum to Sarvam language codes LANGUAGE_MAP = { Language.BN_IN: "bn-IN", Language.GU_IN: "gu-IN", Language.HI_IN: "hi-IN", Language.KN_IN: "kn-IN", Language.ML_IN: "ml-IN", Language.MR_IN: "mr-IN", Language.TA_IN: "ta-IN", Language.TE_IN: "te-IN", Language.PA_IN: "pa-IN", Language.OR_IN: "od-IN", Language.EN_IN: "en-IN", Language.AS_IN: "as-IN", } return resolve_language(language, LANGUAGE_MAP, use_base_code=False) @dataclass(frozen=True) class ModelConfig: """Immutable configuration for a Sarvam STT model. Attributes: supports_prompt: Whether the model accepts prompt parameter. supports_mode: Whether the model accepts mode parameter. supports_language: Whether the model accepts language parameter. default_language: Default language code (None = auto-detect). default_mode: Default mode (None = not applicable). use_translate_endpoint: Whether to use speech_to_text_translate_streaming endpoint. use_translate_method: Whether to use translate() method instead of transcribe(). """ supports_prompt: bool supports_mode: bool supports_language: bool default_language: Optional[str] default_mode: Optional[str] use_translate_endpoint: bool use_translate_method: bool MODEL_CONFIGS: Dict[str, ModelConfig] = { "saarika:v2.5": ModelConfig( supports_prompt=False, supports_mode=False, supports_language=True, default_language="unknown", default_mode=None, use_translate_endpoint=False, use_translate_method=False, ), "saaras:v2.5": ModelConfig( supports_prompt=True, supports_mode=False, supports_language=False, default_language=None, # Auto-detects language default_mode=None, use_translate_endpoint=True, use_translate_method=True, ), "saaras:v3": ModelConfig( supports_prompt=False, supports_mode=True, supports_language=True, default_language="unknown", default_mode="transcribe", use_translate_endpoint=False, use_translate_method=False, ), } @dataclass class SarvamSTTSettings(STTSettings): """Settings for SarvamSTTService. Parameters: prompt: Optional prompt to guide transcription/translation style/context. Only applicable to models that support prompts (e.g., saaras:v2.5). vad_signals: Enable VAD signals in response. high_vad_sensitivity: Enable high VAD sensitivity. """ prompt: str | None | _NotGiven = field(default_factory=lambda: NOT_GIVEN) vad_signals: bool | None | _NotGiven = field(default_factory=lambda: NOT_GIVEN) high_vad_sensitivity: bool | None | _NotGiven = field(default_factory=lambda: NOT_GIVEN) class SarvamSTTService(STTService): """Sarvam speech-to-text service. Provides real-time speech recognition using Sarvam's WebSocket API. Event handlers available (in addition to STTService events): - on_connected(service): Connected to Sarvam WebSocket - on_disconnected(service): Disconnected from Sarvam WebSocket - on_connection_error(service, error): Connection error occurred Example:: @stt.event_handler("on_connected") async def on_connected(service): ... """ Settings = SarvamSTTSettings _settings: SarvamSTTSettings class InputParams(BaseModel): """Configuration parameters for Sarvam STT service. .. deprecated:: 0.0.105 Use ``settings=SarvamSTTSettings(...)`` instead. Parameters: language: Target language for transcription. - saarika:v2.5: Defaults to "unknown" (auto-detect supported) - saaras:v2.5: Not used (auto-detects language) - saaras:v3: Defaults to "unknown" (auto-detect supported) prompt: Optional prompt to guide transcription/translation style/context. Only applicable to saaras:v2.5. Defaults to None. mode: Mode of operation for saaras:v3 models only. Options: transcribe, translate, verbatim, translit, codemix. Defaults to "transcribe" for saaras:v3. vad_signals: Enable VAD signals in response. Defaults to None. high_vad_sensitivity: Enable high VAD (Voice Activity Detection) sensitivity. Defaults to None. """ language: Optional[Language] = None prompt: Optional[str] = None mode: Optional[Literal["transcribe", "translate", "verbatim", "translit", "codemix"]] = None vad_signals: Optional[bool] = None high_vad_sensitivity: Optional[bool] = None def __init__( self, *, api_key: str, model: Optional[str] = None, mode: Optional[ Literal["transcribe", "translate", "verbatim", "translit", "codemix"] ] = None, sample_rate: Optional[int] = None, input_audio_codec: str = "wav", params: Optional[InputParams] = None, settings: Optional[SarvamSTTSettings] = None, ttfs_p99_latency: Optional[float] = SARVAM_TTFS_P99, keepalive_timeout: Optional[float] = None, keepalive_interval: float = 5.0, **kwargs, ): """Initialize the Sarvam STT service. Args: api_key: Sarvam API key for authentication. model: Sarvam model to use for transcription. .. deprecated:: 0.0.105 Use ``settings=SarvamSTTSettings(model=...)`` instead. mode: Mode of operation. Options: transcribe, translate, verbatim, translit, codemix. Only applicable to models that support it (e.g., saaras:v3). Defaults to the model's default mode. sample_rate: Audio sample rate. Defaults to 16000 if not specified. input_audio_codec: Audio codec/format of the input file. Defaults to "wav". params: Configuration parameters for Sarvam STT service. .. deprecated:: 0.0.105 Use ``settings=SarvamSTTSettings(...)`` instead. settings: Runtime-updatable settings. When provided alongside deprecated parameters, ``settings`` values take precedence. ttfs_p99_latency: P99 latency from speech end to final transcript in seconds. Override for your deployment. See https://github.com/pipecat-ai/stt-benchmark keepalive_timeout: Seconds of no audio before sending silence to keep the connection alive. None disables keepalive. keepalive_interval: Seconds between idle checks when keepalive is enabled. **kwargs: Additional arguments passed to the parent STTService. """ # --- 1. Hardcoded defaults --- default_settings = SarvamSTTSettings( model="saarika:v2.5", language=None, prompt=None, vad_signals=None, high_vad_sensitivity=None, ) # --- 2. Deprecated direct-arg overrides --- if model is not None: _warn_deprecated_param("model", SarvamSTTSettings, "model") default_settings.model = model # --- 3. Deprecated params overrides --- if params is not None: _warn_deprecated_param("params", SarvamSTTSettings) if not settings: default_settings.language = params.language default_settings.prompt = params.prompt if params.mode is not None: mode = params.mode default_settings.vad_signals = params.vad_signals default_settings.high_vad_sensitivity = params.high_vad_sensitivity # --- 4. Settings delta (canonical API, always wins) --- if settings is not None: default_settings.apply_update(settings) # Resolve model config and validate (after all overrides) resolved_model = default_settings.model if resolved_model not in MODEL_CONFIGS: allowed = ", ".join(sorted(MODEL_CONFIGS.keys())) raise ValueError(f"Unsupported model '{resolved_model}'. Allowed values: {allowed}.") self._config = MODEL_CONFIGS[resolved_model] # Validate parameters against model capabilities if default_settings.prompt is not None and not self._config.supports_prompt: raise ValueError(f"Model '{resolved_model}' does not support prompt parameter.") if mode is not None and not self._config.supports_mode: raise ValueError(f"Model '{resolved_model}' does not support mode parameter.") if default_settings.language is not None and not self._config.supports_language: raise ValueError( f"Model '{resolved_model}' does not support language parameter (auto-detects language)." ) # Resolve mode default from model config if mode is None: mode = self._config.default_mode super().__init__( sample_rate=sample_rate, ttfs_p99_latency=ttfs_p99_latency, keepalive_timeout=keepalive_timeout, keepalive_interval=keepalive_interval, settings=default_settings, **kwargs, ) self._api_key = api_key # Init-only connection config (not runtime-updatable) self._mode = mode # Store connection parameters self._input_audio_codec = input_audio_codec # Initialize Sarvam SDK client self._sdk_headers = sdk_headers() # Pass Pipecat SDK headers directly at client construction time so they are # merged by the Sarvam SDK's client wrapper and consistently applied to # WebSocket handshake requests. self._sarvam_client = AsyncSarvamAI(api_subscription_key=api_key, headers=self._sdk_headers) self._websocket_context = None self._socket_client = None self._receive_task = None if default_settings.vad_signals: self._register_event_handler("on_speech_started") self._register_event_handler("on_speech_stopped") self._register_event_handler("on_utterance_end") logger.info(f"Sarvam STT initialized with SDK headers: {self._sdk_headers}") def language_to_service_language(self, language: Language) -> str: """Convert pipecat Language enum to Sarvam's language code. Args: language: The Language enum value to convert. Returns: The Sarvam language code string. """ return language_to_sarvam_language(language) def _get_language_string(self) -> Optional[str]: """Resolve the current language setting to a Sarvam language code string.""" if self._settings.language: return language_to_sarvam_language(self._settings.language) return self._config.default_language def can_generate_metrics(self) -> bool: """Check if this service can generate processing metrics. Returns: True, as Sarvam service supports metrics generation. """ return True async def process_frame(self, frame: Frame, direction: FrameDirection): """Process incoming frames. Handles VAD frames for TTFB tracking when using Pipecat's VAD instead of Sarvam's built-in VAD. """ await super().process_frame(frame, direction) # Only handle VAD frames when not using Sarvam's VAD signals if not self._settings.vad_signals: if isinstance(frame, VADUserStartedSpeakingFrame): await self._start_metrics() elif isinstance(frame, VADUserStoppedSpeakingFrame): if self._socket_client: await self._socket_client.flush() async def _update_settings(self, delta: STTSettings) -> dict[str, Any]: """Apply a settings delta, validate, sync state, and reconnect. Args: delta: A :class:`STTSettings` (or ``SarvamSTTSettings``) delta. Returns: Dict mapping changed field names to their previous values. Raises: ValueError: If a setting is not supported by the current model. """ # Validate against model capabilities before applying if is_given(delta.language) and delta.language is not None: if not self._config.supports_language: raise ValueError( f"Model '{self._settings.model}' does not support language parameter " "(auto-detects language)." ) if ( isinstance(delta, SarvamSTTSettings) and is_given(delta.prompt) and delta.prompt is not None ): if not self._config.supports_prompt: raise ValueError( f"Model '{self._settings.model}' does not support prompt parameter." ) changed = await super()._update_settings(delta) # Language and prompt are WebSocket connect-time parameters; reconnect to apply. reconnect_fields = {"language", "prompt"} if changed.keys() & reconnect_fields: await self._disconnect() await self._connect() unhandled = {k: v for k, v in changed.items() if k not in reconnect_fields} if unhandled: self._warn_unhandled_updated_settings(unhandled) return changed async def set_prompt(self, prompt: Optional[str]): """Set the transcription/translation prompt and reconnect. .. deprecated:: 0.0.104 Use ``STTUpdateSettingsFrame(SarvamSTTSettings(prompt=...))`` instead. Args: prompt: Prompt text to guide transcription/translation style/context. Pass None to clear/disable prompt. Only applicable to models that support prompts. """ import warnings with warnings.catch_warnings(): warnings.simplefilter("always") warnings.warn( f"{self.__class__.__name__}.set_prompt() is deprecated. " "Use STTUpdateSettingsFrame(SarvamSTTSettings(prompt=...)) instead.", DeprecationWarning, stacklevel=2, ) if not self._config.supports_prompt: if prompt is not None: raise ValueError( f"Model '{self._settings.model}' does not support prompt parameter." ) # If prompt is None and model doesn't support prompts, silently return (no-op) return logger.info(f"Updating {self._settings.model} prompt.") self._settings.prompt = prompt await self._disconnect() await self._connect() async def start(self, frame: StartFrame): """Start the Sarvam STT service. Args: frame: The start frame containing initialization parameters. """ await super().start(frame) await self._connect() async def stop(self, frame: EndFrame): """Stop the Sarvam STT service. Args: frame: The end frame. """ await super().stop(frame) await self._disconnect() async def cancel(self, frame: CancelFrame): """Cancel the Sarvam STT service. Args: frame: The cancel frame. """ await super().cancel(frame) await self._disconnect() async def run_stt(self, audio: bytes) -> AsyncGenerator[Frame, None]: """Send audio data to Sarvam for transcription. Args: audio: Raw audio bytes to transcribe. Yields: Frame: None (transcription results come via WebSocket callbacks). """ if not self._socket_client: yield None return try: # Convert audio bytes to base64 for Sarvam API audio_base64 = base64.b64encode(audio).decode("utf-8") # Convert input_audio_codec to encoding format (prepend "audio/" if needed) encoding = ( self._input_audio_codec if self._input_audio_codec.startswith("audio/") else f"audio/{self._input_audio_codec}" ) # Build method arguments method_kwargs = { "audio": audio_base64, "encoding": encoding, "sample_rate": self.sample_rate, } # Use appropriate method based on model configuration if self._config.use_translate_method: await self._socket_client.translate(**method_kwargs) else: await self._socket_client.transcribe(**method_kwargs) except Exception as e: yield ErrorFrame(error=f"Error sending audio to Sarvam: {e}", exception=e) yield None async def _connect(self): """Connect to Sarvam WebSocket API using the SDK.""" logger.debug("Connecting to Sarvam") try: # Build common connection parameters connect_kwargs = { "model": self._settings.model, "sample_rate": str(self.sample_rate), } # Enable flush signal when using Pipecat's VAD (not Sarvam's) so that # the flush() call on user-stopped-speaking is honored by the server. if not self._settings.vad_signals: connect_kwargs["flush_signal"] = "true" # Only send vad parameters when explicitly set (avoid overriding server defaults) if self._settings.vad_signals is not None: connect_kwargs["vad_signals"] = "true" if self._settings.vad_signals else "false" if self._settings.high_vad_sensitivity is not None: connect_kwargs["high_vad_sensitivity"] = ( "true" if self._settings.high_vad_sensitivity else "false" ) # Add language_code for models that support it language_string = self._get_language_string() if language_string is not None: connect_kwargs["language_code"] = language_string # Add mode for models that support it if self._config.supports_mode and self._mode is not None: connect_kwargs["mode"] = self._mode # Prompt support differs across sarvamai versions. Prefer connect-time prompt # when available and gracefully degrade if the SDK doesn't accept it. if self._settings.prompt is not None and self._config.supports_prompt: connect_kwargs["prompt"] = self._settings.prompt def _connect_with_sdk_headers(connect_fn, **kwargs): # If prompt is unsupported at connect-time, retry without it. # Headers are supplied through request_options because this is a # documented SDK parameter that survives SDK signature changes. request_options = {"additional_headers": self._sdk_headers} attempts = [kwargs] if "prompt" in kwargs: attempts.append({k: v for k, v in kwargs.items() if k != "prompt"}) last_type_error = None for attempt_kwargs in attempts: try: return connect_fn( **attempt_kwargs, request_options=request_options, ) except TypeError as e: last_type_error = e try: # Fallback for SDK builds that don't expose request_options. return connect_fn(**attempt_kwargs) except TypeError as e: last_type_error = e if last_type_error is not None: raise last_type_error return connect_fn(**kwargs) # Choose the appropriate endpoint based on model configuration if self._config.use_translate_endpoint: self._websocket_context = _connect_with_sdk_headers( self._sarvam_client.speech_to_text_translate_streaming.connect, **connect_kwargs, ) else: self._websocket_context = _connect_with_sdk_headers( self._sarvam_client.speech_to_text_streaming.connect, **connect_kwargs, ) # Enter the async context manager self._socket_client = await self._websocket_context.__aenter__() # Fallback for SDKs that support runtime prompt updates. if self._settings.prompt is not None and self._config.supports_prompt: prompt_setter = getattr(self._socket_client, "set_prompt", None) if callable(prompt_setter): await prompt_setter(self._settings.prompt) # Register event handler for incoming messages def _message_handler(message): """Wrapper to handle async response handler.""" # Use Pipecat's built-in task management self.create_task(self._handle_message(message)) self._socket_client.on(EventType.MESSAGE, _message_handler) # Start receive task using Pipecat's task management self._receive_task = self.create_task(self._receive_task_handler()) self._create_keepalive_task() logger.info("Connected to Sarvam successfully") except ApiError as e: self._socket_client = None self._websocket_context = None await self.push_error(error_msg=f"Sarvam API error: {e}", exception=e) except Exception as e: self._socket_client = None self._websocket_context = None await self.push_error(error_msg=f"Failed to connect to Sarvam: {e}", exception=e) async def _disconnect(self): """Disconnect from Sarvam WebSocket API using SDK.""" await self._cancel_keepalive_task() if self._receive_task: await self.cancel_task(self._receive_task) self._receive_task = None # Clear references first to prevent run_stt from sending audio # during the close handshake. socket_client = self._socket_client websocket_context = self._websocket_context self._socket_client = None self._websocket_context = None if websocket_context and socket_client: try: await websocket_context.__aexit__(None, None, None) except Exception as e: await self.push_error( error_msg=f"Error closing WebSocket connection: {e}", exception=e ) finally: logger.debug("Disconnected from Sarvam WebSocket") async def _receive_task_handler(self): """Handle incoming messages from Sarvam WebSocket. This task wraps the SDK's start_listening() method which processes messages via the registered event handler callback. """ if not self._socket_client: return try: # Start listening for messages from the Sarvam SDK # Messages will be handled via the _message_handler callback await self._socket_client.start_listening() except Exception as e: await self.push_error(error_msg=f"Sarvam receive task error: {e}", exception=e) async def _handle_message(self, message): """Handle incoming WebSocket message from Sarvam SDK. Processes transcription data and VAD events from the Sarvam service. Args: message: The parsed response object from Sarvam WebSocket. """ logger.debug(f"Received response: {message}") try: if message.type == "events": # VAD event signal = message.data.signal_type timestamp = message.data.occured_at logger.debug(f"VAD Signal: {signal}, Occurred at: {timestamp}") if signal == "START_SPEECH": await self._start_metrics() logger.debug("User started speaking") await self._call_event_handler("on_speech_started") await self.broadcast_frame(UserStartedSpeakingFrame) await self.broadcast_interruption() elif signal == "END_SPEECH": logger.debug("User stopped speaking") await self._call_event_handler("on_speech_stopped") await self.broadcast_frame(UserStoppedSpeakingFrame) elif message.type == "data": transcript = message.data.transcript language_code = message.data.language_code # Prefer language from message (auto-detected for translate models). Fallback to configured. if language_code: language = self._map_language_code_to_enum(language_code) else: language_string = self._get_language_string() if language_string: language = self._map_language_code_to_enum(language_string) else: language = Language.HI_IN # Emit utterance end event await self._call_event_handler("on_utterance_end") if transcript and transcript.strip(): # Record tracing for this transcription event await self._handle_transcription(transcript, True, language) await self.push_frame( TranscriptionFrame( transcript, self._user_id, time_now_iso8601(), language, result=(message.dict() if hasattr(message, "dict") else str(message)), ) ) await self.stop_processing_metrics() except Exception as e: await self.push_error(error_msg=f"Failed to handle message: {e}", exception=e) await self.stop_all_metrics() @traced_stt async def _handle_transcription( self, transcript: str, is_final: bool, language: Optional[Language] = None ): """Handle a transcription result with tracing. This method is decorated with @traced_stt for observability. """ pass def _map_language_code_to_enum(self, language_code: str) -> Language: """Map Sarvam language code to pipecat Language enum.""" mapping = { "bn-IN": Language.BN_IN, "gu-IN": Language.GU_IN, "hi-IN": Language.HI_IN, "kn-IN": Language.KN_IN, "ml-IN": Language.ML_IN, "mr-IN": Language.MR_IN, "ta-IN": Language.TA_IN, "te-IN": Language.TE_IN, "pa-IN": Language.PA_IN, "od-IN": Language.OR_IN, "en-US": Language.EN_US, "en-IN": Language.EN_IN, "as-IN": Language.AS_IN, } return mapping.get(language_code, Language.HI_IN) def _is_keepalive_ready(self) -> bool: """Check if the Sarvam SDK websocket client is connected.""" return self._socket_client is not None async def _send_keepalive(self, silence: bytes): """Send silent audio via the Sarvam SDK to keep the connection alive. Args: silence: Silent 16-bit mono PCM audio bytes. """ audio_base64 = base64.b64encode(silence).decode("utf-8") encoding = ( self._input_audio_codec if self._input_audio_codec.startswith("audio/") else f"audio/{self._input_audio_codec}" ) method_kwargs = { "audio": audio_base64, "encoding": encoding, "sample_rate": self.sample_rate, } if self._config.use_translate_method: await self._socket_client.translate(**method_kwargs) else: await self._socket_client.transcribe(**method_kwargs) async def _start_metrics(self): """Start processing metrics collection.""" await self.start_processing_metrics()