WeSpeaker Python Binding

This is a python binding of WeSpeaker.

WeSpeaker is a production first and production ready end-to-end speaker recognition toolkit.

1. Two onnx models are available: voxceleb model, cnceleb_model
2. Extract embedding from wav file or feature(Fbank/MFCC).
3. Support using kaldiio to save embedding.

Install

Python 3.6+ is required.

pip3 install wespeakerruntime

Usage

Extract embedding from wav file

import sys
import wespeakerruntime as wespeaker
wav_file = sys.argv[1]
speaker = wespeaker.Speaker(lang='chs')
ans = speaker.extract_embedding(wav_file)
print(ans)

You can also specify the following parameter in wespeaker.Speaker

• onnx_path (str, optional): is the path of onnx model.
  • Default: onnx model will be downloaded from the server.
• lang (str): chs for cnceleb_model. en for voxceleb model.
• inter_op_num_threads and intra_op_num_threads (int): the number of threads during the model runing. For details, please see: https://onnxruntime.ai/docs/

The parameters of extract_embedding

• wav_path (str): the path of wav
• resample_rate (int): resampling rate. Default: 16000
• num_mel_bins (int): dimension of fbank. Default: 80
• frame_length (int): frame length. Default: 25
• frame_shift (int): frame shift. Default: 10
• cmn (bool): if true, cepstrum average normalization is applied. Default: True

Compute cosine score

import wespeakerruntime as wespeaker
speaker = wespeaker.Speaker(lang='chs')
emb1 = speaker.extract_embedding_wav(wav1_path)
emb2 = speaker.extract_embedding_wav(wav2_path)
score = speaker.compute_cosine_score(emb1, emb2)

The parameters of compute_cosine_score:

• emb1(numpy.ndarray): embedding of speaker-1
• emb2(numpy.ndarray): embedding of speaker-2

[Optional] Extract embedding from feature(fbank/MFCC)

import sys
import wespeakerruntime as wespeaker
feat = your_fbank
speaker = wespeaker.Speaker(lang='chs')
ans = speaker.extract_embedding_feat(feat)
print(ans)

The parameters of extract_embedding_feat:

• feats(numpy.ndarray): the shape is [B, T, D].
• cmn(bool): if true, cepstrum average normalization is applied. Default: True

[Optional] Extract embedding from wav.scp

import sys
import wespeakerruntime as wespeaker
wav_scp = sys.argv[1]
speaker = wespeaker.Speaker(lang='chs')
speaker.extract_embedding_kaldiio(wav_scp, 'embed.ark')

The parameters of extract_embedding_kaldiio:

• wav_path (str): the path of wav
• embed_ark (str): the path of $ouput.ark
• resample_rate (int): resampling rate. Default: 16000
• num_mel_bins(int): dimension of fbank. Default: 80
• frame_length(int): frame length. Default: 25
• frame_shift(int): frame shift. Default: 10
• cmn(bool): if true, cepstrum average normalization is applied. Default: True

Build on Your Local Machine

git clone git@github.com:wenet-e2e/wespeaker.git
cd wespeaker/runtime/binding/python
python setup.py install