Repository files navigation

This repository provides a pre-trained model for extracting thex-vector (speaker representation vector). The model is trained usingJTubeSpeech corpus, a Japanese speech corpus collected from YouTube.

このリポジトリは，x-vector (話者表現ベクトル) を抽出するための学習済みモデルを提供します．このモデルは，JTubeSpeechコーパスと呼ばれる，YouTubeから収集した日本語音声から学習されています．

Instantiate the pre-trained model without explicit install as follow:

importtorchmodel=torch.hub.load("sarulab-speech/xvector_jtubespeech","xvector",trust_repo=True)

Then, follow 'Usage / 使い方' section.

• The number of speakers: 1,233
• Sampling frequency: 16,000Hz
• Speaker recognition accuracy: 91% (test data)
• Feature: 24-dimensional MFCC
• Dimensionality of x-vector: 512
• Other configurations: followed the ASV recipe for VoxCeleb in Kaldi.
  • In the opensourced model, model parameters of recognition layers following to the x-vector layer were randomized to protect data privacy.

pip install xvector-jtubespeech

import numpy as npfrom scipy.io import wavfileimport torchfrom torchaudio.compliance import kaldifrom xvector_jtubespeech import XVectordef extract_xvector(  model, # xvector model  wav   # 16kHz mono):  # extract mfcc  wav = torch.from_numpy(wav.astype(np.float32)).unsqueeze(0)  mfcc = kaldi.mfcc(wav, num_ceps=24, num_mel_bins=24) # [1, T, 24]  mfcc = mfcc.unsqueeze(0)  # extract xvector  xvector = model.vectorize(mfcc) # (1, 512)  xvector = xvector.to("cpu").detach().numpy().copy()[0]    return xvector_, wav = wavfile.read("sample.wav") # 16kHz monomodel = XVector("xvector.pth")xvector = extract_xvector(model, wav) # (512, )

• Takaki Hamada / 濱田 誉輝 (The University of Tokyo / 東京大学)
• Shinnosuke Takamichi / 高道 慎之介 (The University of Tokyo / 東京大学)

MIT

• The audio samplesample.wav was copied fromPJS corpus.