fix(transcription): parse cue timings only for subtitle responses

This commit is contained in:
见安 2026-10-04 19:18:52 +08:00
parent 7e66a3a141
commit 1babfc6524
4 changed files with 53 additions and 18 deletions

View file

@ -6,6 +6,7 @@ from pydantic import BaseModel
if TYPE_CHECKING:
from aiohttp import ClientSession
from litellm.litellm_core_utils.audio_utils.subtitle_utils import SUBTITLE_RESPONSE_FORMATS
from litellm.litellm_core_utils.audio_utils.utils import get_audio_file_name
from litellm.litellm_core_utils.core_helpers import set_provider_response_headers_in_hidden_params
from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
@ -128,7 +129,11 @@ class OpenAIAudioTranscription(OpenAIChatCompletion):
if isinstance(response, BaseModel):
stringified_response = response.model_dump()
else:
duration: Final = extract_duration_from_srt_or_vtt(response)
duration: Final = (
extract_duration_from_srt_or_vtt(response)
if data.get("response_format") in SUBTITLE_RESPONSE_FORMATS
else None
)
stringified_response = {
**TranscriptionResponse(text=response).model_dump(),
**({"_audio_transcription_duration": duration} if duration is not None else {}),

View file

@ -10126,25 +10126,24 @@ def extract_duration_from_srt_or_vtt(srt_or_vtt_content: str) -> float | None:
srt_or_vtt_content (str): The content of an SRT or VTT file as a string.
Returns:
Optional[float]: The total duration in seconds, or None if no timestamps are found.
Optional[float]: The total duration in seconds, or None if no cue timestamps are found.
"""
# Regular expression to match timestamps in the format "hh:mm:ss,ms" or "hh:mm:ss.ms"
timestamp_pattern: Final = r"(\d{2}):(\d{2}):(\d{2})[.,](\d{3})"
timestamp_pattern: Final = (
r"^[ \t]*(?:\d{2,}:)?\d{2}:\d{2}[.,]\d{3}[ \t]+-->[ \t]+"
r"(?:(\d{2,}):)?(\d{2}):(\d{2})[.,](\d{3})"
)
timestamps: Final[Sequence[tuple[str, str, str, str]]] = re.findall(timestamp_pattern, srt_or_vtt_content)
timestamps: Final[Sequence[tuple[str, str, str, str]]] = re.findall(
timestamp_pattern, srt_or_vtt_content, re.MULTILINE
)
if not timestamps:
return None
# Convert timestamps to seconds and find the max (end time)
durations: Final = []
match: tuple[str, str, str, str]
for match in timestamps:
hours, minutes, seconds, milliseconds = map(int, match)
total_seconds = hours * 3600 + minutes * 60 + seconds + milliseconds / 1000.0
durations.append(total_seconds)
return max(durations) if durations else None
return max(
(
int(hours or 0) * 3600 + int(minutes) * 60 + int(seconds) + int(milliseconds) / 1000.0
for hours, minutes, seconds, milliseconds in timestamps
),
default=None,
)
def _add_path_to_api_base(api_base: str, ending_path: str) -> str:

View file

@ -68,8 +68,11 @@ def test_audio_transcriptions_records_provider_response_headers():
("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),
("text", "The timecode is 00:00:12.500", None),
("text", "00:00:01.000 --> 00:00:12.500\nHello world", None),
("vtt", "WEBVTT\n\n00:01.500 --> 00:02.000\nHello world\n", 2.0),
),
ids=("srt", "vtt", "empty-subtitle", "plain-text"),
ids=("srt", "vtt", "empty-subtitle", "plain-text", "text-timecode", "text-cue", "short-vtt"),
)
def test_sync_transcription_retains_text_and_charges_subtitle_duration(
response_format: Literal["srt", "vtt", "text"], subtitle: str, expected_duration: float | None

View file

@ -76,6 +76,7 @@ from litellm.utils import (
async_post_call_success_deployment_hook,
calculate_max_parallel_requests,
client,
extract_duration_from_srt_or_vtt,
get_non_default_completion_params,
get_optional_params_image_gen,
get_prompt_cache_min_tokens,
@ -86,6 +87,33 @@ from litellm.utils import (
# Adds the parent directory to the system path
@pytest.mark.parametrize(
("subtitle", "expected_duration"),
(
("1\n01:01:01,000 --> 01:01:12,500\nHello\n", 3672.5),
("WEBVTT\n\n00:01.500 --> 00:02.000\nHello\n", 2.0),
("WEBVTT\n\n59:59.000 --> 01:00:02.500 align:start\nHello\n", 3602.5),
("WEBVTT\n\n100:00:00.000 --> 100:00:02.000\nHello\n", 360002.0),
("1\n00:00:00,000 --> 00:00:03,000\nThe timecode is 99:00:00.000\n", 3.0),
("WEBVTT\n\n00:00.000 --> 00:05.000\nHello\n\n00:03.000 --> 00:04.000\nWorld\n", 5.0),
("", None),
("The timecode is 00:00:12.500", None),
),
ids=(
"srt-hours",
"vtt-no-hours",
"vtt-mixed-hours",
"vtt-long-hours",
"payload-timecode",
"overlap",
"empty",
"text",
),
)
def test_subtitle_duration_uses_the_latest_cue_end_time(subtitle: str, expected_duration: float | None) -> None:
assert extract_duration_from_srt_or_vtt(subtitle) == expected_duration
def test_non_ocr_wrapper_preserves_logging_executor_and_context(monkeypatch: pytest.MonkeyPatch) -> None:
marker: Final = contextvars.ContextVar("non-ocr-logging-context", default="missing")
token: Final = marker.set("caller-context")