Fast voice activity detection with Python
Voice Activity Detection with Python
pip install vader
import vader # use your own mono, preferably 16kHz .wav file filename = "audio.wav" # returns segments of vocal activity (unit: seconds) # note: it uses a pre-trained logistic regression by default segments = vader.vad(filename) # where to dump audio files out_folder = "segments" # write segments into .wav files vader.vad_to_files(segments, filename, out_folder)
You can also use different pre-trained models by specifying the method parameter
# logistic method segments = vader.vad(filename, threshold=.1, window=20, method="logistic") # multi-layer perceptron method segments = vader.vad(filename, threshold=.1, window=20, method="nn") # Naive Bayes method segments = vader.vad(filename, threshold=.5, window=10, method="nb")
threshold parameter is the ratio of voice frames above which a window of frames is counted as a voiced sample. The
window parameter controls the number of frames considered, and thus the length of the voiced samples.
You can also train your own models:
import vader model = vader.train.logistic_regression(mfccs, activities) model = vader.train.random_forest_classifier(mfccs, activities) model = vader.train.NN(mfccs, activities) model = vader.train.NB(mfccs, activities)
mfccs is a list of varying length mfcc features (num_samples, varying_lengths, 13), while
activities is a list of binary vectors whose lengths match those of the mfcc features (num_samples, varying_lengths), equal to 1 when a frame is voiced, and 0 otherwise.
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, size||File type||Python version||Upload date||Hashes|
|Filename, size vader-0.0.2-py3-none-any.whl (45.9 kB)||File type Wheel||Python version py3||Upload date||Hashes View|
|Filename, size vader-0.0.2.tar.gz (46.7 kB)||File type Source||Python version None||Upload date||Hashes View|