diff --git a/litellm/llms/soniox/realtime/transformation.py b/litellm/llms/soniox/realtime/transformation.py index 2a80371e7e5..805e24f3026 100644 --- a/litellm/llms/soniox/realtime/transformation.py +++ b/litellm/llms/soniox/realtime/transformation.py @@ -1,5 +1,6 @@ import asyncio import time +from collections import deque from collections.abc import Awaitable, Callable, Mapping from itertools import chain, groupby from types import MappingProxyType, TracebackType @@ -143,8 +144,10 @@ def _sample_rate(update: TranscriptionSessionUpdate | None) -> int: return audio_format.rate -def _language_hints(language: str | None, configured: tuple[str, ...] | None) -> tuple[str, ...] | None: - if not language: +def _language_hints( + language: str | None, configured: tuple[str, ...] | None, strict: bool | None +) -> tuple[str, ...] | None: + if not language or (strict and configured): return configured return (language, *(hint for hint in configured or () if hint != language)) @@ -163,7 +166,9 @@ def build_start_request( ) -> SonioxStartRequest: return SonioxStartRequest( **options.model_dump(exclude={"language_hints"}), - language_hints=_language_hints(None if update is None else update.language, options.language_hints), + language_hints=_language_hints( + None if update is None else update.language, options.language_hints, options.language_hints_strict + ), model=model, audio_format="pcm_s16le", sample_rate=_sample_rate(update), @@ -197,7 +202,8 @@ def _changes_started_stream(update: TranscriptionSessionUpdate, started: SonioxS _changes_started_rate(update, started) or ( update.language is not None - and _language_hints(update.language, started.language_hints) != started.language_hints + and _language_hints(update.language, started.language_hints, started.language_hints_strict) + != started.language_hints ) or (turn_detection_set and _uses_endpoint_detection(update, drop_params) != started.enable_endpoint_detection) ) @@ -255,7 +261,7 @@ class SonioxEventTransformer: self._drops_originals: Final = drops_originals self._item_id: str | None = None self._final_tokens: Final[list[_SonioxToken]] = [] # mutable-ok: a turn's tokens append in O(1) each - self._finalizes: tuple[bool, ...] = () + self._finalizes: Final[deque[bool]] = deque() # mutable-ok: pending finalizes queue and pop in O(1) each self._processed_ms: int = 0 self._billed_ms: int = 0 self._refusal: SonioxException | None = None @@ -279,7 +285,7 @@ class SonioxEventTransformer: if discards_turn and not self._finalizes: self._item_id = None self._final_tokens.clear() - self._finalizes = (*self._finalizes, discards_turn) + self._finalizes.append(discards_turn) def take_unbilled_usage(self, stream_ms: int = 0) -> RealtimeInputAudioTranscriptionUsage | None: billable_ms: Final = max(self._processed_ms, stream_ms) @@ -313,11 +319,10 @@ class SonioxEventTransformer: @property def _discarding(self) -> bool: - return self._finalizes[:1] == (True,) + return bool(self._finalizes) and self._finalizes[0] def _finalized(self) -> tuple[OpenAIRealtimeEvents, ...]: - discarded: Final = self._discarding - self._finalizes = self._finalizes[1:] + discarded: Final = self._finalizes.popleft() if self._finalizes else False return () if discarded else self._complete(speech_stopped=False) def _start(self) -> tuple[OpenAIRealtimeEvents, ...]: @@ -479,6 +484,8 @@ class SonioxRealtimeConfig(BaseRealtimeConfig): ) -> tuple[str | bytes, ...]: request: Final = json_object(message, SonioxProtocolError) event_type: Final = request.get("type") + if event_type is None: + raise SonioxProtocolError("Soniox realtime takes OpenAI realtime events, and this message has no type") if event_type in SESSION_UPDATE_EVENT_TYPES: update: Final = parse_transcription_session_update(message, SonioxProtocolError) ignored: Final = _ignored_settings(update) diff --git a/tests/unit/llms/soniox/realtime/test_soniox_realtime_transformation.py b/tests/unit/llms/soniox/realtime/test_soniox_realtime_transformation.py index d52a2e718df..3a6ef57d1ae 100644 --- a/tests/unit/llms/soniox/realtime/test_soniox_realtime_transformation.py +++ b/tests/unit/llms/soniox/realtime/test_soniox_realtime_transformation.py @@ -98,6 +98,23 @@ def test_an_empty_transcription_language_keeps_the_deployment_language_hints(): assert _start_request(frames)["language_hints"] == ["fa"] +def test_a_client_language_cannot_widen_strict_deployment_language_hints(): + config = SonioxRealtimeConfig(options=SonioxRealtimeOptions(language_hints=("fa",), language_hints_strict=True)) + + frames = config.transform_realtime_request(_session_update({"transcription": {"language": "en"}}), MODEL) + + assert _start_request(frames)["language_hints"] == ["fa"] + assert config.transform_realtime_request(_session_update({"transcription": {"language": "en"}}), MODEL) == () + + +def test_a_soniox_native_config_frame_is_rejected_without_starting_the_stream(): + config = SonioxRealtimeConfig() + + with pytest.raises(SonioxProtocolError): + config.transform_realtime_request(json.dumps({"model": "stt-rt-v5", "audio_format": "auto"}), MODEL) + assert _start_request(config.transform_realtime_request(_append(), MODEL))["sample_rate"] == 24_000 + + def test_audio_before_session_update_starts_a_default_stream_once_and_rejects_a_later_change(): config = SonioxRealtimeConfig()