CosyVoice/tools/extract_embedding.py

#!/usr/bin/env python3
# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#   http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import os
from concurrent.futures import ThreadPoolExecutor

import onnxruntime
import torch
import torchaudio
import torchaudio.compliance.kaldi as kaldi
from tqdm import tqdm


def extract_embedding(input_list):
    utt, wav_file, ort_session = input_list

    audio, sample_rate = torchaudio.load(wav_file)
    if sample_rate != 16000:
        audio = torchaudio.transforms.Resample(
            orig_freq=sample_rate, new_freq=16000
        )(audio)
    feat = kaldi.fbank(audio, num_mel_bins=80, dither=0, sample_frequency=16000)
    feat = feat - feat.mean(dim=0, keepdim=True)
    embedding = (
        ort_session.run(
            None,
            {
                ort_session.get_inputs()[0]
                .name: feat.unsqueeze(dim=0)
                .cpu()
                .numpy()
            },
        )[0]
        .flatten()
        .tolist()
    )
    return (utt, embedding)


def main(args):
    utt2wav, utt2spk = {}, {}
    with open("{}/wav.scp".format(args.dir)) as f:
        for l in f:
            l = l.replace("\n", "").split()
            utt2wav[l[0]] = l[1]
    with open("{}/utt2spk".format(args.dir)) as f:
        for l in f:
            l = l.replace("\n", "").split()
            utt2spk[l[0]] = l[1]

    assert os.path.exists(args.onnx_path), "onnx_path not exists"

    option = onnxruntime.SessionOptions()
    option.graph_optimization_level = (
        onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
    )
    option.intra_op_num_threads = 1
    providers = ["CPUExecutionProvider"]
    ort_session = onnxruntime.InferenceSession(
        args.onnx_path, sess_options=option, providers=providers
    )

    inputs = [
        (utt, utt2wav[utt], ort_session)
        for utt in tqdm(utt2wav.keys(), desc="Load data")
    ]
    with ThreadPoolExecutor(max_workers=args.num_thread) as executor:
        results = list(
            tqdm(
                executor.map(extract_embedding, inputs),
                total=len(inputs),
                desc="Process data: ",
            )
        )

    utt2embedding, spk2embedding = {}, {}
    for utt, embedding in results:
        utt2embedding[utt] = embedding
        spk = utt2spk[utt]
        if spk not in spk2embedding:
            spk2embedding[spk] = []
        spk2embedding[spk].append(embedding)

    for k, v in spk2embedding.items():
        spk2embedding[k] = torch.tensor(v).mean(dim=0).tolist()

    torch.save(utt2embedding, "{}/utt2embedding.pt".format(args.dir))
    torch.save(spk2embedding, "{}/spk2embedding.pt".format(args.dir))


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--dir", type=str)
    parser.add_argument("--onnx_path", type=str)
    parser.add_argument("--num_thread", type=int, default=8)
    args = parser.parse_args()
    main(args)