fix(transcription): preserve duration for synchronous subtitle responses

This commit is contained in:
见安 2026-10-04 19:03:35 +08:00
parent a2bf67a037
commit 7e66a3a141
2 changed files with 57 additions and 2 deletions

View file

@ -128,7 +128,11 @@ class OpenAIAudioTranscription(OpenAIChatCompletion):
if isinstance(response, BaseModel):
stringified_response = response.model_dump()
else:
stringified_response = TranscriptionResponse(text=response).model_dump()
duration: Final = extract_duration_from_srt_or_vtt(response)
stringified_response = {
**TranscriptionResponse(text=response).model_dump(),
**({"_audio_transcription_duration": duration} if duration is not None else {}),
}
## LOGGING
logging_obj.post_call(

View file

@ -1,10 +1,14 @@
from typing import Final
from datetime import datetime, timezone
from typing import Final, Literal
from unittest.mock import Mock
import httpx
import pytest
from openai import AsyncOpenAI, OpenAI
from litellm import completion_cost
from litellm.litellm_core_utils.litellm_logging import Logging
from litellm.llms.openai.cost_calculation import cost_per_second
from litellm.llms.openai.transcriptions.handler import OpenAIAudioTranscription
from litellm.types.utils import TranscriptionResponse
@ -57,6 +61,53 @@ def test_audio_transcriptions_records_provider_response_headers():
_assert_headers_recorded(response, logging_obj)
@pytest.mark.parametrize(
("response_format", "subtitle", "expected_duration"),
(
("srt", "1\n00:00:00,000 --> 00:00:01,000\nHello\n\n2\n00:00:01,000 --> 00:00:12,500\nWorld\n", 12.5),
("vtt", "WEBVTT\n\n00:00:00.000 --> 00:00:01.000\nHello\n\n00:00:01.000 --> 00:00:12.500\nWorld\n", 12.5),
("srt", "", None),
("text", "Hello world", None),
),
ids=("srt", "vtt", "empty-subtitle", "plain-text"),
)
def test_sync_transcription_retains_text_and_charges_subtitle_duration(
response_format: Literal["srt", "vtt", "text"], subtitle: str, expected_duration: float | None
) -> None:
logging_obj: Final = Logging(
model="whisper-1",
messages=None,
stream=False,
call_type="transcription",
start_time=datetime(2025, 1, 1, tzinfo=timezone.utc),
litellm_call_id="subtitle-transcription",
function_id="subtitle-transcription",
)
transport: Final = httpx.MockTransport(
lambda request: httpx.Response(200, text=subtitle, headers={"content-type": "text/plain"})
)
with httpx.Client(transport=transport) as http_client:
response: Final = OpenAIAudioTranscription().audio_transcriptions(
model="whisper-1",
audio_file=("audio.wav", b"riff-bytes", "audio/wav"),
optional_params={"response_format": response_format},
litellm_params={},
model_response=TranscriptionResponse(),
timeout=10.0,
max_retries=0,
logging_obj=logging_obj,
api_key="transport-only",
api_base=None,
client=OpenAI(api_key="transport-only", http_client=http_client),
)
assert response.model_dump() == TranscriptionResponse(text=subtitle).model_dump()
assert response._hidden_params.get("audio_transcription_duration") == expected_duration
assert completion_cost(completion_response=response, model="whisper-1", call_type="transcription") == sum(
cost_per_second(model="whisper-1", custom_llm_provider="openai", duration=expected_duration or 0.0)
)
@pytest.mark.asyncio
async def test_async_audio_transcriptions_records_provider_response_headers():
logging_obj = _logging_obj()