aishell example

游雁

2024-02-19 94de39dde2e616a01683c518023d0fab72b4e103

 funasr/utils/prepare_data.py

@@ -5,9 +5,9 @@

import kaldiio
import numpy as np
import librosa
import torch.distributed as dist
import torchaudio
import soundfile


def filter_wav_text(data_dir, dataset):
@@ -46,7 +46,7 @@
    try:
        waveform, sampling_rate = torchaudio.load(wav_path)
    except:
        waveform, sampling_rate = soundfile.read(wav_path)
        waveform, sampling_rate = librosa.load(wav_path)
        waveform = np.expand_dims(waveform, axis=0)
    n_frames = (waveform.shape[1] * 1000.0) / (sampling_rate * frontend_conf["frame_shift"] * frontend_conf["lfr_n"])
    feature_dim = frontend_conf["n_mels"] * frontend_conf["lfr_m"]
@@ -87,6 +87,7 @@
                sample_name, feature_path = line.strip().split()
                feature = kaldiio.load_mat(feature_path)
                n_frames, feature_dim = feature.shape
                write_flag = True
                if n_frames > 0 and length_min > 0:
                    write_flag = n_frames >= length_min
                if n_frames > 0 and length_max > 0:
@@ -198,6 +199,7 @@
    data_names = args.dataset_conf.get("data_names", "speech,text").split(",")
    data_types = args.dataset_conf.get("data_types", "sound,text").split(",")
    file_names = args.data_file_names.split(",")
    batch_type = args.dataset_conf["batch_conf"]["batch_type"]
    print("data_names: {}, data_types: {}, file_names: {}".format(data_names, data_types, file_names))
    assert len(data_names) == len(data_types) == len(file_names)
    if args.dataset_type == "small":
@@ -228,7 +230,7 @@
            filter_wav_text(args.data_dir, args.train_set)
            filter_wav_text(args.data_dir, args.valid_set)

        if args.dataset_type == "small":
        if args.dataset_type == "small" and batch_type != "unsorted":
            calc_shape(args, args.train_set)
            calc_shape(args, args.valid_set)