594 lines
22 KiB
Python
594 lines
22 KiB
Python
import numpy as np
|
|
import math, os, csv
|
|
import torchaudio
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.utils.data as data
|
|
import torch.distributed as dist
|
|
import soundfile as sf
|
|
from torch.utils.data import Dataset
|
|
import torch.utils.data as data
|
|
import os
|
|
import sys
|
|
sys.path.append(os.path.dirname(__file__))
|
|
from pydub import AudioSegment
|
|
from .misc import read_and_config_file, get_file_extension
|
|
import librosa
|
|
import random
|
|
EPS = 1e-6
|
|
MAX_WAV_VALUE_16B = 32768.0
|
|
MAX_WAV_VALUE_32B = 2147483648.0
|
|
|
|
def audioread_archieved(path, sampling_rate):
|
|
"""
|
|
Reads an audio file from the specified path, normalizes the audio,
|
|
resamples it to the desired sampling rate (if necessary), and ensures it is single-channel.
|
|
|
|
Parameters:
|
|
path (str): The file path of the audio file to be read.
|
|
sampling_rate (int): The target sampling rate for the audio.
|
|
|
|
Returns:
|
|
numpy.ndarray: The processed audio data, normalized, resampled (if necessary),
|
|
and converted to mono (if the input audio has multiple channels).
|
|
"""
|
|
|
|
# Read audio data and its sample rate from the file.
|
|
data, fs = sf.read(path)
|
|
|
|
# convert to mono channel
|
|
if len(data.shape) >1:
|
|
if data.shape[0] > data.shape[1]:
|
|
data = data[:, 0]
|
|
else:
|
|
data = data[0, :]
|
|
|
|
# Normalize the audio data.
|
|
data, scalar = audio_norm(data)
|
|
|
|
# Resample the audio if the sample rate is different from the target sampling rate.
|
|
if fs != sampling_rate:
|
|
data = librosa.resample(data, orig_sr=fs, target_sr=sampling_rate)
|
|
|
|
# Convert to mono by selecting the first channel if the audio has multiple channels.
|
|
if len(data.shape) > 1:
|
|
data = data[:, 0]
|
|
|
|
# Return the processed audio data.
|
|
return data, scalar
|
|
|
|
def read_audio(file_path):
|
|
"""
|
|
Use AudioSegment to load audio from all supported audio input format
|
|
"""
|
|
|
|
try:
|
|
audio = AudioSegment.from_file(file_path)
|
|
return audio
|
|
except Exception as e:
|
|
print(f"Error loading file: {e}")
|
|
return None
|
|
|
|
def audioread(path, sampling_rate, use_norm):
|
|
"""
|
|
Reads an audio file from the specified path, normalizes the audio,
|
|
resamples it to the desired sampling rate (if necessary), and ensures it is single-channel.
|
|
|
|
Parameters:
|
|
path (str): The file path of the audio file to be read.
|
|
sampling_rate (int): The target sampling rate for the audio.
|
|
use_norm (bool): The flag for specifying whether using input audio normalization
|
|
|
|
Returns:
|
|
numpy.ndarray: The processed audio data, normalized, resampled (if necessary),
|
|
and converted to mono (if the input audio has multiple channels).
|
|
"""
|
|
|
|
# Read audio data and its sample rate from the file.
|
|
audio_info = {}
|
|
ext = get_file_extension(path).replace('.', '')
|
|
audio_info['ext']=ext
|
|
|
|
try:
|
|
data = AudioSegment.from_file(path)
|
|
except Exception as e:
|
|
print(f"Error loading file: {e}")
|
|
return None
|
|
|
|
data = read_audio(path)
|
|
|
|
audio_info['sample_rate'] = data.frame_rate
|
|
audio_info['channels'] = data.channels
|
|
audio_info['sample_width'] = data.sample_width
|
|
|
|
data_array = np.array(data.get_array_of_samples())
|
|
if max(data_array) > MAX_WAV_VALUE_16B:
|
|
audio_np = data_array / MAX_WAV_VALUE_32B
|
|
else:
|
|
audio_np = data_array / MAX_WAV_VALUE_16B
|
|
|
|
audios = []
|
|
# Check if the audio is stereo
|
|
if audio_info['channels'] == 2:
|
|
audios.append(audio_np[::2]) # Even indices (left channel)
|
|
audios.append(audio_np[1::2]) # Odd indices (right channel)
|
|
else:
|
|
audios.append(audio_np)
|
|
|
|
# Normalize the audio data.
|
|
audios_normed = []
|
|
scalars = []
|
|
for audio in audios:
|
|
if use_norm:
|
|
audio_normed, scalar = audio_norm(audio)
|
|
audios_normed.append(audio_normed)
|
|
scalars.append(scalar)
|
|
else:
|
|
audios_normed.append(audio)
|
|
scalars.append(1)
|
|
# Resample the audio if the sample rate is different from the target sampling rate.
|
|
if audio_info['sample_rate'] != sampling_rate:
|
|
index = 0
|
|
for audio_normed in audios_normed:
|
|
audios_normed[index] = librosa.resample(audio_normed, orig_sr=audio_info['sample_rate'], target_sr=sampling_rate)
|
|
index = index + 1
|
|
|
|
# Return the processed audio data.
|
|
return audios_normed, scalars, audio_info
|
|
|
|
def audio_norm(x):
|
|
"""
|
|
Normalizes the input audio signal to a target Root Mean Square (RMS) level,
|
|
applying two stages of scaling. This ensures the audio signal is neither too quiet
|
|
nor too loud, keeping its amplitude consistent.
|
|
|
|
Parameters:
|
|
x (numpy.ndarray): Input audio signal to be normalized.
|
|
|
|
Returns:
|
|
numpy.ndarray: Normalized audio signal.
|
|
"""
|
|
|
|
# Compute the root mean square (RMS) of the input audio signal.
|
|
rms = (x ** 2).mean() ** 0.5
|
|
|
|
# Calculate the scalar to adjust the signal to the target level (-25 dB).
|
|
scalar = 10 ** (-25 / 20) / (rms + EPS)
|
|
|
|
# Scale the input audio by the computed scalar.
|
|
x = x * scalar
|
|
|
|
# Compute the power of the scaled audio signal.
|
|
pow_x = x ** 2
|
|
|
|
# Calculate the average power of the audio signal.
|
|
avg_pow_x = pow_x.mean()
|
|
|
|
# Compute RMS only for audio segments with higher-than-average power.
|
|
rmsx = pow_x[pow_x > avg_pow_x].mean() ** 0.5
|
|
|
|
# Calculate another scalar to further normalize based on higher-power segments.
|
|
scalarx = 10 ** (-25 / 20) / (rmsx + EPS)
|
|
|
|
# Apply the second scalar to the audio.
|
|
x = x * scalarx
|
|
|
|
# Return the doubly normalized audio signal.
|
|
return x, 1/(scalar * scalarx + EPS)
|
|
|
|
class DataReader(object):
|
|
"""
|
|
A class for reading audio data from a list of files, normalizing it,
|
|
and extracting features for further processing. It supports extracting
|
|
features from each file, reshaping the data, and returning metadata
|
|
like utterance ID and data length.
|
|
|
|
Parameters:
|
|
args: Arguments containing the input path and target sampling rate.
|
|
|
|
Attributes:
|
|
file_list (list): A list of audio file paths to process.
|
|
sampling_rate (int): The target sampling rate for audio files.
|
|
"""
|
|
|
|
def __init__(self, args):
|
|
# Read and configure the file list from the input path provided in the arguments.
|
|
# The file list is decoded, if necessary.
|
|
self.file_list = read_and_config_file(args, args.input_path, decode=True)
|
|
|
|
# Store the target sampling rate.
|
|
self.sampling_rate = args.sampling_rate
|
|
|
|
# Store the args file
|
|
self.args = args
|
|
|
|
def __len__(self):
|
|
"""
|
|
Returns the number of audio files in the file list.
|
|
|
|
Returns:
|
|
int: Number of files to process.
|
|
"""
|
|
return len(self.file_list)
|
|
|
|
def __getitem__(self, index):
|
|
"""
|
|
Retrieves the features of the audio file at the given index.
|
|
|
|
Parameters:
|
|
index (int): Index of the file in the file list.
|
|
|
|
Returns:
|
|
tuple: Features (inputs, utterance ID, data length) for the selected audio file.
|
|
"""
|
|
if self.args.task == 'target_speaker_extraction':
|
|
if self.args.network_reference.cue== 'lip':
|
|
return self.file_list[index]
|
|
return self.extract_feature(self.file_list[index])
|
|
|
|
def extract_feature(self, path):
|
|
"""
|
|
Extracts features from the given audio file path.
|
|
|
|
Parameters:
|
|
path (str): The file path of the audio file.
|
|
|
|
Returns:
|
|
inputs (numpy.ndarray): Reshaped audio data for further processing.
|
|
utt_id (str): The unique identifier of the audio file, usually the filename.
|
|
length (int): The length of the original audio data.
|
|
"""
|
|
# Extract the utterance ID from the file path (usually the filename).
|
|
utt_id = path.split('/')[-1]
|
|
use_norm = False
|
|
|
|
#We suggest to use norm for 'FRCRN_SE_16K' and 'MossFormer2_SS_16K' models
|
|
if self.args.network in ['FRCRN_SE_16K','MossFormer2_SS_16K'] :
|
|
use_norm = True
|
|
|
|
# Read and normalize the audio data, converting it to float32 for processing.
|
|
audios_norm, scalars, audio_info = audioread(path, self.sampling_rate, use_norm)
|
|
|
|
if self.args.network in ['MossFormer2_SR_48K']:
|
|
audio_info['sample_rate'] = self.sampling_rate
|
|
|
|
for i in range(len(audios_norm)):
|
|
audios_norm[i] = audios_norm[i].astype(np.float32)
|
|
# Reshape the data to ensure it's in the format [1, data_length].
|
|
audios_norm[i] = np.reshape(audios_norm[i], [1, audios_norm[i].shape[0]])
|
|
|
|
# Return the reshaped audio data, utterance ID, and the length of the original data.
|
|
return audios_norm, utt_id, audios_norm[0].shape[1], scalars, audio_info
|
|
|
|
class Wave_Processor(object):
|
|
"""
|
|
A class for processing audio data, specifically for reading input and label audio files,
|
|
segmenting them into fixed-length segments, and applying padding or trimming as necessary.
|
|
|
|
Methods:
|
|
process(path, segment_length, sampling_rate):
|
|
Processes audio data by reading, padding, or segmenting it to match the specified segment length.
|
|
|
|
Parameters:
|
|
path (dict): A dictionary containing file paths for 'inputs' and 'labels' audio files.
|
|
segment_length (int): The desired length of audio segments to extract.
|
|
sampling_rate (int): The target sampling rate for reading the audio files.
|
|
"""
|
|
|
|
def process(self, path, segment_length, sampling_rate):
|
|
"""
|
|
Reads input and label audio files, and ensures the audio is segmented into
|
|
the desired length, padding if necessary or extracting random segments if
|
|
the audio is longer than the target segment length.
|
|
|
|
Parameters:
|
|
path (dict): Dictionary containing the paths to 'inputs' and 'labels' audio files.
|
|
segment_length (int): Desired length of the audio segment in samples.
|
|
sampling_rate (int): Target sample rate for the audio.
|
|
|
|
Returns:
|
|
tuple: A pair of numpy arrays representing the processed input and label audio,
|
|
either padded to the segment length or trimmed.
|
|
"""
|
|
# Read the input and label audio files using the target sampling rate.
|
|
wave_inputs = audioread(path['inputs'], sampling_rate)
|
|
wave_labels = audioread(path['labels'], sampling_rate)
|
|
|
|
# Get the length of the label audio (assumed both inputs and labels have similar lengths).
|
|
len_wav = wave_labels.shape[0]
|
|
|
|
# If the input audio is shorter than the desired segment length, pad it with zeros.
|
|
if wave_inputs.shape[0] < segment_length:
|
|
# Create zero-padded arrays for inputs and labels.
|
|
padded_inputs = np.zeros(segment_length, dtype=np.float32)
|
|
padded_labels = np.zeros(segment_length, dtype=np.float32)
|
|
|
|
# Copy the original audio into the padded arrays.
|
|
padded_inputs[:wave_inputs.shape[0]] = wave_inputs
|
|
padded_labels[:wave_labels.shape[0]] = wave_labels
|
|
else:
|
|
# Randomly select a start index for segmenting the audio if it's longer than the segment length.
|
|
st_idx = random.randint(0, len_wav - segment_length)
|
|
|
|
# Extract a segment of the desired length from the inputs and labels.
|
|
padded_inputs = wave_inputs[st_idx:st_idx + segment_length]
|
|
padded_labels = wave_labels[st_idx:st_idx + segment_length]
|
|
|
|
# Return the processed (padded or segmented) input and label audio.
|
|
return padded_inputs, padded_labels
|
|
|
|
class Fbank_Processor(object):
|
|
"""
|
|
A class for processing input audio data into mel-filterbank (Fbank) features,
|
|
including the computation of delta and delta-delta features.
|
|
|
|
Methods:
|
|
process(inputs, args):
|
|
Processes the raw audio input and returns the mel-filterbank features
|
|
along with delta and delta-delta features.
|
|
"""
|
|
|
|
def process(self, inputs, args):
|
|
# Convert frame length and shift from seconds to milliseconds.
|
|
frame_length = int(args.win_len / args.sampling_rate * 1000)
|
|
frame_shift = int(args.win_inc / args.sampling_rate * 1000)
|
|
|
|
# Set up configuration for the mel-filterbank computation.
|
|
fbank_config = {
|
|
"dither": 1.0,
|
|
"frame_length": frame_length,
|
|
"frame_shift": frame_shift,
|
|
"num_mel_bins": args.num_mels,
|
|
"sample_frequency": args.sampling_rate,
|
|
"window_type": args.win_type
|
|
}
|
|
|
|
# Convert the input audio to a FloatTensor and scale it to match the expected input range.
|
|
inputs = torch.FloatTensor(inputs * MAX_WAV_VALUE)
|
|
|
|
# Compute the mel-filterbank features using Kaldi's fbank function.
|
|
fbank = torchaudio.compliance.kaldi.fbank(inputs.unsqueeze(0), **fbank_config)
|
|
|
|
# Add delta and delta-delta features.
|
|
fbank_tr = torch.transpose(fbank, 0, 1)
|
|
fbank_delta = torchaudio.functional.compute_deltas(fbank_tr)
|
|
fbank_delta_delta = torchaudio.functional.compute_deltas(fbank_delta)
|
|
fbank_delta = torch.transpose(fbank_delta, 0, 1)
|
|
fbank_delta_delta = torch.transpose(fbank_delta_delta, 0, 1)
|
|
|
|
# Concatenate the original Fbank, delta, and delta-delta features.
|
|
fbanks = torch.cat([fbank, fbank_delta, fbank_delta_delta], dim=1)
|
|
|
|
return fbanks.numpy()
|
|
|
|
class AudioDataset(Dataset):
|
|
"""
|
|
A dataset class for loading and processing audio data from different data types
|
|
(train, validation, test). Supports audio processing and feature extraction
|
|
(e.g., waveform processing, Fbank feature extraction).
|
|
|
|
Parameters:
|
|
args: Arguments containing dataset configuration (paths, sampling rate, etc.).
|
|
data_type (str): The type of data to load (train, val, test).
|
|
"""
|
|
|
|
def __init__(self, args, data_type):
|
|
self.args = args
|
|
self.sampling_rate = args.sampling_rate
|
|
|
|
# Read the list of audio files based on the data type.
|
|
if data_type == 'train':
|
|
self.wav_list = read_and_config_file(args.tr_list)
|
|
elif data_type == 'val':
|
|
self.wav_list = read_and_config_file(args.cv_list)
|
|
elif data_type == 'test':
|
|
self.wav_list = read_and_config_file(args.tt_list)
|
|
else:
|
|
print(f'Data type: {data_type} is unknown!')
|
|
|
|
# Initialize processors for waveform and Fbank features.
|
|
self.wav_processor = Wave_Processor()
|
|
self.fbank_processor = Fbank_Processor()
|
|
|
|
# Clip data to a fixed segment length based on the sampling rate and max length.
|
|
self.segment_length = self.sampling_rate * self.args.max_length
|
|
print(f'No. {data_type} files: {len(self.wav_list)}')
|
|
|
|
def __len__(self):
|
|
# Return the number of audio files in the dataset.
|
|
return len(self.wav_list)
|
|
|
|
def __getitem__(self, index):
|
|
# Get the input and label paths from the list.
|
|
data_info = self.wav_list[index]
|
|
|
|
# Process the waveform inputs and labels.
|
|
inputs, labels = self.wav_processor.process(
|
|
{'inputs': data_info['inputs'], 'labels': data_info['labels']},
|
|
self.segment_length,
|
|
self.sampling_rate
|
|
)
|
|
|
|
# Optionally load Fbank features if specified.
|
|
if self.args.load_fbank is not None:
|
|
fbanks = self.fbank_processor.process(inputs, self.args)
|
|
return inputs * MAX_WAV_VALUE, labels * MAX_WAV_VALUE, fbanks
|
|
|
|
return inputs, labels
|
|
|
|
def zero_pad_concat(self, inputs):
|
|
"""
|
|
Concatenates a list of input arrays, applying zero-padding as needed to ensure
|
|
they all match the length of the longest input.
|
|
|
|
Parameters:
|
|
inputs (list of numpy arrays): List of input arrays to be concatenated.
|
|
|
|
Returns:
|
|
numpy.ndarray: A zero-padded array with concatenated inputs.
|
|
"""
|
|
|
|
# Get the maximum length among all inputs.
|
|
max_t = max(inp.shape[0] for inp in inputs)
|
|
|
|
# Determine the shape of the output based on the input dimensions.
|
|
shape = None
|
|
if len(inputs[0].shape) == 1:
|
|
shape = (len(inputs), max_t)
|
|
elif len(inputs[0].shape) == 2:
|
|
shape = (len(inputs), max_t, inputs[0].shape[1])
|
|
|
|
# Initialize an array with zeros to hold the concatenated inputs.
|
|
input_mat = np.zeros(shape, dtype=np.float32)
|
|
|
|
# Copy the input data into the zero-padded array.
|
|
for e, inp in enumerate(inputs):
|
|
if len(inp.shape) == 1:
|
|
input_mat[e, :inp.shape[0]] = inp
|
|
elif len(inp.shape) == 2:
|
|
input_mat[e, :inp.shape[0], :] = inp
|
|
|
|
return input_mat
|
|
|
|
def collate_fn_2x_wavs(data):
|
|
"""
|
|
A custom collate function for combining batches of waveform input and label pairs.
|
|
|
|
Parameters:
|
|
data (list): List of tuples (inputs, labels).
|
|
|
|
Returns:
|
|
tuple: Batched inputs and labels as torch.FloatTensors.
|
|
"""
|
|
inputs, labels = zip(*data)
|
|
x = torch.FloatTensor(inputs)
|
|
y = torch.FloatTensor(labels)
|
|
return x, y
|
|
|
|
def collate_fn_2x_wavs_fbank(data):
|
|
"""
|
|
A custom collate function for combining batches of waveform inputs, labels, and Fbank features.
|
|
|
|
Parameters:
|
|
data (list): List of tuples (inputs, labels, fbanks).
|
|
|
|
Returns:
|
|
tuple: Batched inputs, labels, and Fbank features as torch.FloatTensors.
|
|
"""
|
|
inputs, labels, fbanks = zip(*data)
|
|
x = torch.FloatTensor(inputs)
|
|
y = torch.FloatTensor(labels)
|
|
z = torch.FloatTensor(fbanks)
|
|
return x, y, z
|
|
|
|
class DistributedSampler(data.Sampler):
|
|
"""
|
|
Sampler for distributed training. Divides the dataset among multiple replicas (processes),
|
|
ensuring that each process gets a unique subset of the data. It also supports shuffling
|
|
and managing epochs.
|
|
|
|
Parameters:
|
|
dataset (Dataset): The dataset to sample from.
|
|
num_replicas (int): Number of processes participating in the training.
|
|
rank (int): Rank of the current process.
|
|
shuffle (bool): Whether to shuffle the data or not.
|
|
seed (int): Random seed for reproducibility.
|
|
"""
|
|
|
|
def __init__(self, dataset, num_replicas=None, rank=None, shuffle=True, seed=0):
|
|
if num_replicas is None:
|
|
if not dist.is_available():
|
|
raise RuntimeError("Requires distributed package to be available")
|
|
num_replicas = dist.get_world_size()
|
|
if rank is None:
|
|
if not dist.is_available():
|
|
raise RuntimeError("Requires distributed package to be available")
|
|
rank = dist.get_rank()
|
|
|
|
self.dataset = dataset
|
|
self.num_replicas = num_replicas
|
|
self.rank = rank
|
|
self.epoch = 0
|
|
self.num_samples = int(math.ceil(len(self.dataset) * 1.0 / self.num_replicas))
|
|
self.total_size = self.num_samples * self.num_replicas
|
|
self.shuffle = shuffle
|
|
self.seed = seed
|
|
|
|
def __iter__(self):
|
|
# Shuffle the indices based on the epoch and seed.
|
|
if self.shuffle:
|
|
g = torch.Generator()
|
|
g.manual_seed(self.seed + self.epoch)
|
|
ind = torch.randperm(int(len(self.dataset) / self.num_replicas), generator=g) * self.num_replicas
|
|
indices = []
|
|
for i in range(self.num_replicas):
|
|
indices = indices + (ind + i).tolist()
|
|
else:
|
|
indices = list(range(len(self.dataset)))
|
|
|
|
# Add extra samples to make the dataset evenly divisible.
|
|
indices += indices[:(self.total_size - len(indices))]
|
|
assert len(indices) == self.total_size
|
|
|
|
# Subsample for the current process.
|
|
indices = indices[self.rank * self.num_samples:(self.rank + 1) * self.num_samples]
|
|
assert len(indices) == self.num_samples
|
|
|
|
return iter(indices)
|
|
|
|
def __len__(self):
|
|
return self.num_samples
|
|
|
|
def set_epoch(self, epoch):
|
|
self.epoch = epoch
|
|
|
|
def get_dataloader(args, data_type):
|
|
"""
|
|
Creates and returns a data loader and sampler for the specified dataset type (train, validation, or test).
|
|
|
|
Parameters:
|
|
args (Namespace): Configuration arguments containing details such as batch size, sampling rate,
|
|
network type, and whether distributed training is used.
|
|
data_type (str): The type of dataset to load ('train', 'val', 'test').
|
|
|
|
Returns:
|
|
sampler (DistributedSampler or None): The sampler for distributed training, or None if not used.
|
|
generator (DataLoader): The PyTorch DataLoader for the specified dataset.
|
|
"""
|
|
|
|
# Initialize the dataset based on the given arguments and dataset type (train, val, or test).
|
|
datasets = AudioDataset(args=args, data_type=data_type)
|
|
|
|
# Create a distributed sampler if distributed training is enabled; otherwise, use no sampler.
|
|
sampler = DistributedSampler(
|
|
datasets,
|
|
num_replicas=args.world_size, # Number of replicas in distributed training.
|
|
rank=args.local_rank # Rank of the current process.
|
|
) if args.distributed else None
|
|
|
|
# Select the appropriate collate function based on the network type.
|
|
if args.network == 'FRCRN_SE_16K' or args.network == 'MossFormerGAN_SE_16K':
|
|
# Use the collate function for two-channel waveform data (inputs and labels).
|
|
collate_fn = collate_fn_2x_wavs
|
|
elif args.network == 'MossFormer2_SE_48K':
|
|
# Use the collate function for waveforms along with Fbank features.
|
|
collate_fn = collate_fn_2x_wavs_fbank
|
|
else:
|
|
# Print an error message if the network type is unknown.
|
|
print(f'in dataloader, please specify a correct network type using args.network!')
|
|
return
|
|
|
|
# Create a DataLoader with the specified dataset, batch size, and worker configuration.
|
|
generator = data.DataLoader(
|
|
datasets,
|
|
batch_size=args.batch_size, # Batch size for training.
|
|
shuffle=(sampler is None), # Shuffle the data only if no sampler is used.
|
|
collate_fn=collate_fn, # Use the selected collate function for batching data.
|
|
num_workers=args.num_workers, # Number of workers for data loading.
|
|
sampler=sampler # Use the distributed sampler if applicable.
|
|
)
|
|
|
|
# Return both the sampler and DataLoader (generator).
|
|
return sampler, generator
|
|
|