examples/wenet/infer.py

#!/usr/bin/python3
# -*- coding: utf-8 -*-
import argparse
import os
from pathlib import Path
import sys

pwd = os.path.abspath(os.path.dirname(__file__))
sys.path.append(os.path.join(pwd, "../../"))

import librosa
import numpy as np
import sherpa
from scipy.io import wavfile
import torch
import torchaudio

from project_settings import project_path, temp_directory


def get_args():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--model_dir",
        default=(project_path / "pretrained_models/huggingface/csukuangfj/wenet-chinese-model").as_posix(),
        type=str
    )
    parser.add_argument(
        "--filename",
        default=(project_path / "data/test_wavs/paraformer-zh/si_chuan_hua.wav").as_posix(),
        type=str
    )
    parser.add_argument("--sample_rate", default=16000, type=int)
    args = parser.parse_args()
    return args


def main():
    args = get_args()

    model_dir = Path(args.model_dir)
    nn_model_file = model_dir / "final.zip"
    tokens_file = model_dir / "units.txt"

    print("nn_model_file: {}".format(nn_model_file))
    print("tokens_file: {}".format(tokens_file))

    feat_config = sherpa.FeatureConfig(normalize_samples=False)
    feat_config.fbank_opts.frame_opts.samp_freq = args.sample_rate
    feat_config.fbank_opts.mel_opts.num_bins = 80
    feat_config.fbank_opts.frame_opts.dither = 0

    config = sherpa.OfflineRecognizerConfig(
        nn_model=nn_model_file.as_posix(),
        tokens=tokens_file.as_posix(),
        use_gpu=False,
        feat_config=feat_config,
        decoding_method="greedy_search",
        num_active_paths=2,
    )
    recognizer = sherpa.OfflineRecognizer(config)

    signal, sample_rate = librosa.load(args.filename, sr=args.sample_rate)
    signal *= 32768.0
    signal = np.array(signal, dtype=np.int16)

    temp_file = temp_directory / "temp.wav"
    wavfile.write(
        temp_file.as_posix(),
        rate=args.sample_rate,
        data=signal
    )

    s = recognizer.create_stream()

    s.accept_wave_file(
        temp_file.as_posix()
    )
    recognizer.decode_stream(s)

    text = s.result.text.strip()
    text = text.lower()
    print("text: {}".format(text))
    return


if __name__ == "__main__":
    main()