dguad-vad 0.1.0

No project description provided

Dguard VAD

Based on Silero VAD and RnNoise.

Installation

pip install dguard_vad

or

git clone http://ai.lyxxkj.com.cn:3001/zhaosheng/dguard_vad.git
cd dguard_vad
pip install -e .

The model files will not be downloaded automatically. You need to download them manually and put them in the right place.(Default: $$DGUARD_MODEL_PATH/dguard_vad.onnx)

$DGUARD_MODEL_PATH is an environment variable that used in all dgurad* projects.

Usage

1. VAD class

from dguard_vad import VAD
SR = 16000
WAV_PATH = "../data/test_16k.wav"
vad = VAD(SR)
# Use get_speech_timestamps to get 
# start and end timestamps of speech segments
timestamps = vad.get_speech_timestamps(WAV_PATH)
for _ in timestamps:
    print(_)

# You shuold get the following output:
# {'segment': 0, 'start': 26560, 'end': 48704}
# {'segment': 1, 'start': 71616, 'end': 106048}
# {'segment': 2, 'start': 149952, 'end': 185920}

# Use get_speech_probs to get probabilities for each chunk
probs = vad.get_speech_probs(WAV_PATH)
for _ in probs:
    print(_)
# You shuold get the following output:
# 0.02
# 0.01
# 0.01
# 0.01
# 0.0

2. VAD class with noise suppression You just need to set denoise=True when initializing VAD class.

vad = VAD(SR, denoise=True)

Please note that: func:get_speech_probs may not work well with noise suppression.

3. VADIterator class Please refer to ws_app/ws_server.py for more details.