[MM] Allow audio chunking for offline LLM (#34628)

Signed-off-by: NickLucche <nlucches@redhat.com>
This commit is contained in:
Nicolò Lucchesi
2026-02-23 21:04:28 -08:00
committed by GitHub
parent 33a0d43c71
commit f91808ae0d
5 changed files with 366 additions and 54 deletions
+4 -1
View File
@@ -33,4 +33,7 @@ class SpeechToTextConfig:
@property
def allow_audio_chunking(self) -> bool:
return self.min_energy_split_window_size is not None
return (
self.min_energy_split_window_size is not None
and self.max_audio_clip_s is not None
)
@@ -45,6 +45,7 @@ from vllm.model_executor.models import (
SupportsTranscription,
supports_transcription,
)
from vllm.multimodal.audio import split_audio
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
@@ -323,11 +324,24 @@ class OpenAISpeechToText(OpenAIServing):
self.asr_config.allow_audio_chunking
and duration > self.asr_config.max_audio_clip_s
)
chunks = [y] if not do_split_audio else self._split_audio(y, int(sr))
if not do_split_audio:
chunks = [y]
else:
assert self.asr_config.max_audio_clip_s is not None
assert self.asr_config.min_energy_split_window_size is not None
chunks = split_audio(
audio_data=y,
sample_rate=int(sr),
max_clip_duration_s=self.asr_config.max_audio_clip_s,
overlap_duration_s=self.asr_config.overlap_chunk_second,
min_energy_window_size=self.asr_config.min_energy_split_window_size,
)
if language is None and getattr(
self.model_cls, "supports_explicit_language_detection", False
):
# Auto-detect language from the first chunk.
language = await self._detect_language(
chunks[0], f"{request_id}-lang_detect"
)
@@ -754,55 +768,3 @@ class OpenAISpeechToText(OpenAIServing):
yield f"data: {data}\n\n"
# Send the final done message after all response.n are finished
yield "data: [DONE]\n\n"
def _split_audio(
self, audio_data: np.ndarray, sample_rate: int
) -> list[np.ndarray]:
assert self.asr_config.max_audio_clip_s is not None, (
f"{self.asr_config.max_audio_clip_s=} cannot be None to"
" split audio into chunks."
)
chunk_size = sample_rate * self.asr_config.max_audio_clip_s
overlap_size = sample_rate * self.asr_config.overlap_chunk_second
chunks = []
i = 0
while i < audio_data.shape[-1]:
if i + chunk_size >= audio_data.shape[-1]:
# handle last chunk
chunks.append(audio_data[..., i:])
break
# Find the best split point in the overlap region
search_start = i + chunk_size - overlap_size
search_end = min(i + chunk_size, audio_data.shape[-1])
split_point = self._find_split_point(audio_data, search_start, search_end)
# Extract chunk up to the split point
chunks.append(audio_data[..., i:split_point])
i = split_point
return chunks
def _find_split_point(self, wav: np.ndarray, start_idx: int, end_idx: int) -> int:
"""Find the best point to split audio by
looking for silence or low amplitude.
Args:
wav: Audio tensor [1, T]
start_idx: Start index of search region
end_idx: End index of search region
Returns:
Index of best splitting point
"""
segment = wav[start_idx:end_idx]
# Calculate RMS energy in small windows
min_energy = math.inf
quietest_idx = 0
min_energy_window = self.asr_config.min_energy_split_window_size
assert min_energy_window is not None
for i in range(0, len(segment) - min_energy_window, min_energy_window):
window = segment[i : i + min_energy_window]
energy = (window**2).mean() ** 0.5
if energy < min_energy:
quietest_idx = i + start_idx
min_energy = energy
return quietest_idx
+118
View File
@@ -216,3 +216,121 @@ class AudioResampler:
f"Invalid resampling method: {self.method}. "
"Supported methods are 'librosa' and 'scipy'."
)
# ============================================================
# Audio Chunking / Splitting
# ============================================================
def split_audio(
audio_data: np.ndarray,
sample_rate: int,
max_clip_duration_s: float,
overlap_duration_s: float,
min_energy_window_size: int,
) -> list[np.ndarray]:
"""Split audio into chunks with intelligent split points.
Splits long audio into smaller chunks at low-energy regions to minimize
cutting through speech. Uses overlapping windows to find quiet moments
for splitting.
Args:
audio_data: Audio array to split. Can be 1D (mono) or multi-dimensional.
Splits along the last dimension (time axis).
sample_rate: Sample rate of the audio in Hz.
max_clip_duration_s: Maximum duration of each chunk in seconds.
overlap_duration_s: Overlap duration in seconds between consecutive chunks.
Used to search for optimal split points.
min_energy_window_size: Window size in samples for finding low-energy regions.
Returns:
List of audio chunks. Each chunk is a numpy array with the same shape
as the input except for the last (time) dimension.
Example:
>>> audio = np.random.randn(1040000) # 65 seconds at 16kHz
>>> chunks = split_audio(
... audio_data=audio,
... sample_rate=16000,
... max_clip_duration_s=30.0,
... overlap_duration_s=1.0,
... min_energy_window_size=1600,
... )
>>> len(chunks)
3
"""
chunk_size = int(sample_rate * max_clip_duration_s)
overlap_size = int(sample_rate * overlap_duration_s)
chunks = []
i = 0
while i < audio_data.shape[-1]:
if i + chunk_size >= audio_data.shape[-1]:
# Handle last chunk - take everything remaining
chunks.append(audio_data[..., i:])
break
# Find the best split point in the overlap region
search_start = i + chunk_size - overlap_size
search_end = min(i + chunk_size, audio_data.shape[-1])
split_point = find_split_point(
audio_data, search_start, search_end, min_energy_window_size
)
# Extract chunk up to the split point
chunks.append(audio_data[..., i:split_point])
i = split_point
return chunks
def find_split_point(
wav: np.ndarray,
start_idx: int,
end_idx: int,
min_energy_window: int,
) -> int:
"""Find the best point to split audio by looking for silence or low amplitude.
Searches for the quietest region within a specified range by calculating
RMS energy in sliding windows.
Args:
wav: Audio array. Can be 1D or multi-dimensional.
start_idx: Start index of search region (inclusive).
end_idx: End index of search region (exclusive).
min_energy_window: Window size in samples for energy calculation.
Returns:
Index of the quietest point within the search region. This is the
recommended split point to minimize audio artifacts.
Example:
>>> audio = np.random.randn(32000)
>>> # Insert quiet region
>>> audio[16000:17600] = 0.01
>>> split_idx = find_split_point(
... wav=audio,
... start_idx=0,
... end_idx=32000,
... min_energy_window=1600,
... )
>>> 16000 <= split_idx <= 17600
True
"""
segment = wav[start_idx:end_idx]
# Calculate RMS energy in small windows
min_energy = math.inf
quietest_idx = 0
for i in range(0, len(segment) - min_energy_window, min_energy_window):
window = segment[i : i + min_energy_window]
energy = (window**2).mean() ** 0.5
if energy < min_energy:
quietest_idx = i + start_idx
min_energy = energy
return quietest_idx