Check if model is Gemini 3 Flash

Strictly validate if a string is valid base64 encoded data. Args: s: String to validate Returns: True if the string is

Heuristic to detect server-side error hints in text payloads/reasons.

Synthesize a (FunctionCall, FunctionCallOutput) pair for a progress update. The new FunctionCall carries ``{call_id}{call_id_suffix}``; name/

Mark the playback as started if it hasn't been already. Must be called under ``audio_lock``.

Capture server-assigned request_id once it is available.

Monitor memory usage and kill the process if it exceeds the limit.

Merge two (language, text) segment lists, combining adjacent segments of the same language.

Per-model option fields shared between the HTTP body and the WS query string.

Check whether the model supports mode parameter.

In-place normalization of client event dicts for legacy Azure compatibility. The legacy Azure Realtime API uses "text" for assistant content part

Record a provider-known id for this stream on the current span. Exposed on the `tts_request_run` span as `lk.provider_request_ids` so users

Build endpoint kwargs for old/new telemetry function signatures.

Normalize a raw WS message into a typed :class:`_ReceivedEvent`. Handles the variant message shapes from SLNG-fronted TTS providers (Sarvam,

parse the list of fields required by the style

Parse timestamp info from API response into TimedString objects. Args: timestamp_info: The timestamp info from the API response.

Replace the TTS's aiohttp session with a stub that captures the body of the first ``post()`` call and short-circuits the rest of the request.

Playout audio frames from the queue to the audio source

Make a POST request to the Tavus API with retry logic. Args: endpoint: API endpoint path (without leading slash)

Internal method to make a POST request with retry logic.

Process an audio frame with Krisp noise reduction. This is the method required by the FrameProcessor interface. Args: fr

A frame whose samples are a monotonic ramp, so splits can be checked for alignment.

Report the total audio duration processed by the STT engine.

Reset all pause tracking state.

Fill in defaults for missing keys. When ``turn_detection`` is a streaming turn detector, keys the caller did not provide fall back to the tig

Layer a user override onto the server defaults. A scalar override replaces the whole map (every language resolves through it); a dict

Resolve language, model and mode from overrides or defaults. Returns: Tuple of (language, model, mode). Raises:

Convert a wrapped tool to a FunctionTool or RawFunctionTool with a warning.

Extract language code from Rime Arcana variant strings. Handles both plain variants ("en", "es") and versioned variants ("3-en", "3-es").