Non-default 'decibel_thres' value causes FSMN VAD model crash

[steven8274]

🐛 Bug

If the 'decibel_thres' of the 'FsmnVADStreaming' class is set to values higher the the default value(-50, for example), the model may crash at this line of 'funasr/models/fsmn_vad_streaming/model.py':

cur_decibel = cache["stats"].decibel[t]

the reported error is:

IndexError: list index out of range

This problem persists in versions 1.1.12, 1.1.14 and 1.3.0(maybe all versions from 1.1.12 to 1.3.0.)

To Reproduce

1.Set 'decibel_thres' of FsmnVADStreaming to -50.
2.Use the VAD pipline to segment audio.

Code sample

import soundfile as sf
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
import torch
import numpy as np

vad_inference_pipeline = pipeline(
    task=Tasks.voice_activity_detection,
    model='iic/speech_fsmn_vad_zh-cn-16k-common-pytorch',
    model_revision="v2.0.4",
    disable_update=True,
    device=device,
    decibel_thres=-50
)

wav, _ = sf.read("video_clip2_16k.wav", dtype="int16")
segments_result = vad_inference_pipeline(input=wav.tobytes(), fs=16000)

Expected behavior

Get the right segemnts of audio.

Environment

• OS (Linux):
• FunASR Version (1.1.12, 1.1.14, 1.3.0):
• ModelScope Version (1.19.0):
• PyTorch Version (2.0.1+cu118):
• How you installed funasr (pip):
• Python version: 3.8.20
• GPU (4090)
• CUDA/cuDNN version (cuda12.1):

video_clip2_16k.wav