Close all channels and resolve _done_fut for the current generation. This prevents consumers from hanging indefinitely when a generation is

Map a Sarvam output_audio_codec value to the MIME type the framework decoder expects.

Compute the speaking rate of the audio using the selected method

Book a flight. Args: origin: where to fly from destination: where to fly to

Create a new connection. Returns: The new connection object Raises: NotImplementedError: If no connect callb

Connect to the LiveKit Inference STT WebSocket.

Return the live connection, creating one if needed. Returns ``(connection, acquire_time, reused)`` — matches the ``ConnectionPool`` s

Decode 8-bit G.711 mu-law/A-law samples to little-endian linear 16-bit PCM.

Return True if any known provider diarization flag is truthy.

Push *frames* until both START_OF_SPEECH and END_OF_SPEECH have fired.

Drain and close all the connections queued for closing.

Emit RealtimeModelMetrics for the current generation.

Close current segment, prepare for next.

Ensure HTTP session is available.

Read Sarvam's ``language_probability`` from a response payload. Returns the value as a float when present and numeric. Falls back to ``1.0``

Normalize outputs from graph.astream(..., stream_mode='messages', [subgraphs]). Expected shapes: - (token, meta) - (namespace, (

Degrade gracefully from adaptive interruption to VAD-based interruption. Called when the adaptive interruption detector encounters an unrecov

Returns the index to insert an item by creation time. Iterates in reverse, assuming items are sorted by `created_at`. Finds

Flush AudioByteStream and resampler for the current WAV segment.

Emit completed FunctionToolCalls from the pending buffer.

Get or create the shared credentials resolver instance. This ensures credential caching works across all RealtimeSession instances.

Current load including reserved slots (accepted but not yet launched).

Construct the headers for the Speechify API.

Get or create an LLMModelUsage for the given provider/model combination.

Get the connection pool, creating if needed.

Gracefully recycle the session, preserving conversation state.

Handle connection-level error by failing all active contexts.

Handle response done but not complete, i.e. cancelled, incomplete or failed. For example this method will emit an error if we receive a "fail

Recursively check if any dict in the schema tree is an empty {}.

Check whether a parsed message carries an explicit error indicator.

Recursively check if any dict in the schema tree contains 'oneOf'.

Run load_fnc in executor. Uses signature to call with or without self.