Add runtime support for wespeaker models

[csukuangfj]

Usage

Download test files

wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speaker-recongition-models/zh_cnceleb_resnet34.onnx

wget https://huggingface.co/openspeech/wespeaker-models/resolve/main/test_wavs/00001_spk1.wav
wget https://huggingface.co/openspeech/wespeaker-models/resolve/main/test_wavs/00024_spk1.wav
wget https://huggingface.co/openspeech/wespeaker-models/resolve/main/test_wavs/00010_spk2.wav

Test code

#!/usr/bin/env python3

import sherpa_onnx
import torchaudio
import torch


def load_audio(filename):
    samples, sample_rate = torchaudio.load(filename)
    return samples[0].contiguous().numpy(), sample_rate


def main():
    model = "./zh_cnceleb_resnet34.onnx"
    config = sherpa_onnx.SpeakerEmbeddingExtractorConfig(
        model=model,
        num_threads=1,
        debug=True,
        provider="cpu",
    )
    print(config)
    extractor = sherpa_onnx.SpeakerEmbeddingExtractor(config)

    samples0, sample_rate0 = load_audio("./00001_spk1.wav")
    samples1, sample_rate1 = load_audio("./00024_spk1.wav")
    #  samples1, sample_rate1 = load_audio("./00010_spk2.wav")

    s0 = extractor.create_stream()
    s1 = extractor.create_stream()

    s0.accept_waveform(sample_rate=sample_rate0, waveform=samples0)
    s1.accept_waveform(sample_rate=sample_rate1, waveform=samples1)

    while extractor.is_ready(s0):
        embedding0 = extractor.compute(s0)

    while extractor.is_ready(s1):
        embedding1 = extractor.compute(s1)

    embedding0 = torch.tensor(embedding0)
    embedding1 = torch.tensor(embedding1)
    print(torch.nn.functional.cosine_similarity(embedding0, embedding1, dim=0))


if __name__ == "__main__":
    main()

TODOs

• VAD + speaker embedding model for speaker diarization (who spoke when)
• VAD + speaker embedding model + ASR (who spoke when and what)
• Android Demo
• APIs for other languages, e.g., C/Go/C#, etc.

[csukuangfj]

A demo video for this pull-request can be found at
https://www.bilibili.com/video/BV1nC4y1Y7Kk/

(Note that it is in Chinese.)