fix(soniox): let a server VAD turn answer only the oldest pending commit

This commit is contained in:
mateo-berri 2026-10-29 04:30:32 -07:00
parent cf4dc70aea
commit 446d33df70
2 changed files with 29 additions and 16 deletions

View file

@ -2,7 +2,6 @@ import asyncio
import time
from collections import deque
from collections.abc import Awaitable, Callable, Mapping
from dataclasses import dataclass
from itertools import chain, groupby
from types import MappingProxyType, TracebackType
from typing import Final
@ -261,19 +260,13 @@ def _soniox_error(response: _SonioxResponse) -> SonioxException | None:
)
@dataclass(frozen=True, slots=True)
class _PendingFinalize:
discards_turn: bool
endpoint_turns: int
class SonioxEventTransformer:
def __init__(self, *, drops_originals: bool) -> None:
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: Final[deque[_PendingFinalize]] = deque() # mutable-ok: pending finalizes pop in O(1)
self._endpoint_turns: int = 0
self._finalizes: Final[deque[bool]] = deque() # mutable-ok: pending finalizes queue and pop in O(1) each
self._oldest_commit_answered: bool = False
self._processed_ms: int = 0
self._billed_ms: int = 0
self._refusal: SonioxException | None = None
@ -297,7 +290,7 @@ class SonioxEventTransformer:
if discards_turn and not self._finalizes:
self._item_id = None
self._final_tokens.clear()
self._finalizes.append(_PendingFinalize(discards_turn=discards_turn, endpoint_turns=self._endpoint_turns))
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)
@ -322,7 +315,7 @@ class SonioxEventTransformer:
if self._discarding:
return ()
if token.text == _ENDPOINT_TOKEN:
return self._complete(speech_stopped=True)
return self._end_turn()
started: Final = self._start()
if not token.is_final or (self._drops_originals and token.translation_status == "original"):
return started
@ -331,18 +324,26 @@ class SonioxEventTransformer:
@property
def _discarding(self) -> bool:
return bool(self._finalizes) and self._finalizes[0].discards_turn
return bool(self._finalizes) and self._finalizes[0]
def _finalized(self) -> tuple[OpenAIRealtimeEvents, ...]:
pending: Final = self._finalizes.popleft() if self._finalizes else None
if pending is not None and pending.discards_turn:
discarded: Final = self._finalizes.popleft() if self._finalizes else None
answered: Final = self._oldest_commit_answered
self._oldest_commit_answered = False
if discarded:
return ()
if self._item_id is not None:
return self._complete(speech_stopped=False)
if pending is None or pending.endpoint_turns != self._endpoint_turns:
if discarded is None or answered:
return ()
return (_completed(_new_item_id(), (), self.take_unbilled_usage()),)
def _end_turn(self) -> tuple[OpenAIRealtimeEvents, ...]:
events: Final = self._complete(speech_stopped=True)
if events and self._finalizes:
self._oldest_commit_answered = True
return events
def _start(self) -> tuple[OpenAIRealtimeEvents, ...]:
if self._item_id is not None:
return ()
@ -361,7 +362,6 @@ class SonioxEventTransformer:
tokens: Final = tuple(self._final_tokens)
self._item_id = None
self._final_tokens.clear()
self._endpoint_turns += speech_stopped
stopped: Final[tuple[OpenAIRealtimeEvents, ...]] = (
(speech_event("input_audio_buffer.speech_stopped", item_id),) if speech_stopped else ()
)

View file

@ -319,6 +319,19 @@ def test_a_commit_soniox_answers_after_its_endpoint_already_ended_the_turn_adds_
assert [event["transcript"] for event in events if "transcript" in event] == ["Hello"]
def test_a_server_vad_turn_answers_only_the_oldest_pending_commit():
config = SonioxRealtimeConfig()
config.transform_realtime_request(_append(), MODEL)
_events(config, [_token("Hello")])
config.transform_realtime_request(_event("input_audio_buffer.commit"), MODEL)
config.transform_realtime_request(_append(), MODEL)
config.transform_realtime_request(_event("input_audio_buffer.commit"), MODEL)
events = _events(config, [_token("<end>"), _token("<fin>"), _token("<fin>")])
assert [event["transcript"] for event in events if "transcript" in event] == ["Hello", ""]
def test_a_clear_without_speech_still_answers_nothing():
config = SonioxRealtimeConfig()
config.transform_realtime_request(_append(), MODEL)