[BugFix][Ming-TTS] Improve Seed-TTS WER accuracy (#4859)

Signed-off-by: LHXuuu <xulianhao.xlh@antgroup.com>
This commit is contained in:
LHXuuu authored and GitHub committed 2026-07-03 12:15:03 +02:00
1 parent a8b2bf11d2
commit 991fb7408b
6 files changed
+59 -45

No files matched your search

+4 -33
View File
@@ -35,8 +35,6 @@ from vllm.benchmarks.lib.endpoint_request_func import (
from vllm.logger import init_logger
from vllm.tokenizers import TokenizerLike
logger = init_logger(__name__)
from vllm_omni.benchmarks.audio_continuity import compute_continuity_stats
from vllm_omni.benchmarks.data_modules.daily_omni_dataset import DailyOmniDataset, DailyOmniSampleRequest
from vllm_omni.benchmarks.data_modules.random_multi_modal_dataset import OmniRandomMultiModalDataset
@@ -51,9 +49,9 @@ from vllm_omni.benchmarks.data_modules.sound_effect_dataset import SoundEffectDa
from vllm_omni.benchmarks.data_modules.ttsd_dataset import TTSDDataset
from vllm_omni.metrics import definitions as defs
logger = init_logger(__name__)
_AUDIO_CONTINUITY_THRESHOLD_ENV = "VLLM_OMNI_BENCH_AUDIO_CONTINUITY_THRESHOLD_S"
_AUDIO_SAMPLE_RATE_ENV = "VLLM_OMNI_BENCH_AUDIO_SAMPLE_RATE"
_AUDIO_CHANNELS_ENV = "VLLM_OMNI_BENCH_AUDIO_CHANNELS"
RETURN_STAGE_METRICS_FIELD = "return_stage_metrics"
_IMAGE_STAGE_METRICS_BACKENDS = frozenset({"openai-image-edits-omni"})
_PRINT_STAGE = False
@@ -111,32 +109,6 @@ def _audio_continuity_threshold_s() -> float:
return max(value, 0.0)
def _audio_pcm_format() -> tuple[int, int]:
"""Return ``(sample_rate, channels)`` of the streamed PCM response.
Defaults to 24 kHz mono (Qwen3-TTS / VoxCPM2 / most MOSS-TTS variants).
Override for models with a different codec output format (e.g.
MOSS-TTS-Local-Transformer-v1.5, which streams 48 kHz stereo PCM) via
``VLLM_OMNI_BENCH_AUDIO_SAMPLE_RATE`` / ``VLLM_OMNI_BENCH_AUDIO_CHANNELS`` -
getting this wrong silently skews audio_duration/audio_rtf/underrun stats.
"""
sample_rate = 24000
channels = 1
raw_sr = os.environ.get(_AUDIO_SAMPLE_RATE_ENV)
if raw_sr:
try:
sample_rate = int(raw_sr)
except ValueError:
logger.warning("Invalid %s=%r; using default %d", _AUDIO_SAMPLE_RATE_ENV, raw_sr, sample_rate)
raw_ch = os.environ.get(_AUDIO_CHANNELS_ENV)
if raw_ch:
try:
channels = int(raw_ch)
except ValueError:
logger.warning("Invalid %s=%r; using default %d", _AUDIO_CHANNELS_ENV, raw_ch, channels)
return sample_rate, channels
def _pcm_s16le_to_seed_tts_wer_bytes(
pcm_bytes: bytes,
*,
@@ -1187,9 +1159,8 @@ async def async_request_openai_audio_speech(
output = MixRequestFuncOutput()
output.prompt_len = request_func_input.prompt_len
# PCM format: 16-bit signed; sample_rate/channels are model-dependent
# (see _audio_pcm_format docstring).
sample_rate, channels = _audio_pcm_format()
# PCM format: 16-bit signed; sample_rate/channels are model-dependent.
sample_rate, channels = defs.stream_pcm_format_from_env()
sample_width = 2 # 16-bit = 2 bytes
st = time.perf_counter()