[Frontend] Add automatic language detection for Whisper transcription (#34342)

Signed-off-by: space_check <roman.vuskov@rwth-aachen.de>
Signed-off-by: Roman <45857014+spacecheck@users.noreply.github.com>
Co-authored-by: Nicolò Lucchesi <nicolo.lucchesi@gmail.com>
Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>
This commit is contained in:
Roman
2026-02-21 04:49:41 -08:00
committed by GitHub
co-authored by Nicolò Lucchesi Nicolò Lucchesi
parent 272b535ab3
commit 98b0205c3c
5 changed files with 249 additions and 10 deletions
@@ -41,7 +41,10 @@ from vllm.exceptions import VLLMValidationError
from vllm.inputs import ProcessorInputs
from vllm.logger import init_logger
from vllm.logprobs import FlatLogprobs, Logprob
from vllm.model_executor.models import SupportsTranscription, supports_transcription
from vllm.model_executor.models import (
SupportsTranscription,
supports_transcription,
)
from vllm.outputs import RequestOutput
from vllm.renderers.inputs import DictPrompt, EncoderDecoderDictPrompt
from vllm.renderers.inputs.preprocess import parse_enc_dec_prompt, parse_model_prompt
@@ -242,10 +245,57 @@ class OpenAISpeechToText(OpenAIServing):
model_cls = get_model_cls(self.model_config)
return cast(type[SupportsTranscription], model_cls)
async def _detect_language(
self,
audio_chunk: np.ndarray,
request_id: str,
) -> str:
"""Auto-detect the spoken language from an audio chunk.
Delegates prompt construction and output parsing to the model class
via ``get_language_detection_prompt`` and
``parse_language_detection_output``.
"""
from vllm.sampling_params import SamplingParams
prompt = self.model_cls.get_language_detection_prompt(
audio_chunk,
self.asr_config,
)
allowed_token_ids = self.model_cls.get_language_token_ids(
self.tokenizer,
)
sampling_params = SamplingParams(
max_tokens=1,
temperature=0.0,
allowed_token_ids=allowed_token_ids,
)
result_generator = self.engine_client.generate(
prompt,
sampling_params,
request_id,
)
final_output: RequestOutput
async for final_output in result_generator:
if final_output.finished:
break
token_ids = list(final_output.outputs[0].token_ids)
lang = self.model_cls.parse_language_detection_output(
token_ids,
self.tokenizer,
)
logger.info("Auto-detected language: '%s'", lang)
return lang
async def _preprocess_speech_to_text(
self,
request: SpeechToTextRequest,
audio_data: bytes,
request_id: str,
) -> tuple[list[ProcessorInputs], float]:
# Validate request
language = self.model_cls.validate_language(request.language)
@@ -274,6 +324,15 @@ class OpenAISpeechToText(OpenAIServing):
and duration > self.asr_config.max_audio_clip_s
)
chunks = [y] if not do_split_audio else self._split_audio(y, int(sr))
if language is None and getattr(
self.model_cls, "supports_explicit_language_detection", False
):
language = await self._detect_language(
chunks[0], f"{request_id}-lang_detect"
)
request.language = language
parsed_prompts: list[DictPrompt] = []
for chunk in chunks:
# The model has control over the construction, as long as it
@@ -435,6 +494,7 @@ class OpenAISpeechToText(OpenAIServing):
engine_prompts, duration_s = await self._preprocess_speech_to_text(
request=request,
audio_data=audio_data,
request_id=request_id,
)
except ValueError as e: