update

嘉渊

2023-07-21 d273f7e12693e5b366cbf2ff7d01dde0264b01d9

 funasr/utils/prepare_data.py

@@ -7,6 +7,7 @@
import numpy as np
import torch.distributed as dist
import torchaudio
import soundfile


def filter_wav_text(data_dir, dataset):
@@ -42,7 +43,11 @@


def wav2num_frame(wav_path, frontend_conf):
    waveform, sampling_rate = torchaudio.load(wav_path)
    try:
        waveform, sampling_rate = torchaudio.load(wav_path)
    except:
        waveform, sampling_rate = soundfile.read(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"]
    return n_frames, feature_dim
@@ -191,12 +196,16 @@

def prepare_data(args, distributed_option):
    distributed = distributed_option.distributed
    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"]
    if not distributed or distributed_option.dist_rank == 0:
        if hasattr(args, "filter_input") and args.filter_input:
            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)

@@ -204,9 +213,6 @@
            generate_data_list(args, args.data_dir, args.train_set)
            generate_data_list(args, args.data_dir, args.valid_set)

    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(",")
    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":