"""
Copyright (c) 2019 Microsoft Corporation. All rights reserved.
MIT License
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
"""
import sys
import torch.utils.data as data
import glob
import numpy as np
from multiprocessing import Queue
import zipfile
import io
import json
from simulation import SimpleSimulator, imagesc
from simulation.freq_analysis import stft
import reader
import os
def _utt2seg(data, seg_len, seg_shift):
""" Cut an utterance (MxN matrix) to segments. """
if data.ndim == 1:
data = np.reshape(data, (1, data.size))
dim, n_fr = data.shape
n_seg = int(np.floor((n_fr - seg_len) / seg_shift)) + 1
seg = []
for i in range(n_seg):
start = i * seg_shift
stop = start + seg_len
seg.append(data[:, start:stop])
return seg
class DataBuffer:
"""This is the class that generate data for neural network training and used with dynamic data simulation.
The job is to prepare speech corpus into individual training samples. E.g. feature-label pairs.
It may also need to call speech simulation to simulated distorted multi-channel array data.
This class generates samples and put them in a buffer (a FIFO queue). """
def __init__(self, data_generator, buffer_size=1000, preload_size=100, randomize=True):
self.data_generator = data_generator
self.buffer_size = buffer_size
self.preload_size = preload_size
self.buffer = []
self.randomize = randomize
def get(self):
"""Generate required number of training samples. """
if len(self.buffer) < self.buffer_size: # maintain minimum number of entries in the buffer.
while len(self.buffer) < self.buffer_size + self.preload_size:
tmp_data = self.data_generator.generate()
self.buffer += tmp_data
if self.randomize:
return_idx = np.random.randint(len(self.buffer))
else:
return_idx = 0
data = self.buffer.pop(return_idx)
return data
def get_len(self):
return self.data_generator.get_len()
class SpeechDataset(data.Dataset):
def __init__(self, config):
self.transform=None
self.sequence_mode = config["data_config"]["sequence_mode"]
# load the three types of source data
dir_noise_streams = None
if "dir_noise_paths" in config:
dir_noise_streams = self._load_streams(config["dir_noise_paths"], config['data_path'], is_speech=False)
rir_streams = None
if "rir_paths" in config:
rir_streams = self._load_streams(config["rir_paths"], config['data_path'], is_speech=False, is_rir=True)
source_streams = self._load_streams(config["source_paths"], config['data_path'], is_speech=True)
self.source_stream_sizes = [i.get_number_of_data() for i in source_streams]
if self.sequence_mode:
self.source_stream_cum_sizes = [self.source_stream_sizes[0]]
for i in range(1, len(self.source_stream_sizes)):
self.source_stream_cum_sizes.append(self.source_stream_cum_sizes[-1] + self.source_stream_sizes[i])
if 'simulation_prob' in config['data_config']:
simulation_prob = config['data_config']['simulation_prob']
else:
simulation_prob = 0
generator_config = DataGeneratorSequenceConfig(
use_reverb=config["data_config"]["use_reverb"],
use_noise=config["data_config"]["use_dir_noise"],
snr_range=[config["data_config"]["snr_min"], config["data_config"]["snr_max"]],
n_hour_per_epoch=config["sweep_size"],
sequence_mode=self.sequence_mode,
load_label=config["data_config"]["load_label"],
seglen=config["data_config"]["seg_len"],
segshift=config["data_config"]["seg_shift"],
use_cmn=config["data_config"]["use_cmn"],
simulation_prob=simulation_prob
)
data_generator = DataGeneratorTrain(source_streams, dir_noise_streams, rir_streams, generator_config, DEBUG=False)
if self.sequence_mode:
self.data_buffer = data_generator
else:
self.data_buffer = DataBuffer(data_generator, buffer_size=20000, preload_size=200, randomize=True)
self.sample_len_seconds = config["data_config"]["seg_len"] * 0.01 # default sampling rate: 100Hz
self.stream_idx_for_transform = [0]
def _load_streams(self, source_list, data_path, is_speech=True, is_rir=False):
source_streams = list()
for i in range(len(source_list)):
corpus_type = source_list[i]['type']
corpus_wav_path = data_path+source_list[i]['wav']
label_paths = []
label_names = []
if 'label' in source_list[i]:
label_paths.append(data_path+source_list[i]['label'])
label_names.append('label')
else:
corpus_label_path = None
if 'aux_label' in source_list[i]:
label_paths.append(data_path+source_list[i]['aux_label'])
label_names.append('aux_label')
else:
corpus_label_path = None
print("%s::_load_streams: loading %s from %s..." % (self.__class__.__name__, corpus_type, corpus_wav_path))
curr_stream = reader.stream.gen_stream_from_zip(corpus_wav_path,
label_files=label_paths,
label_names=label_names,
is_speech_corpus=is_speech,
is_rir=is_rir,
get_duration=False,
corpus_name=corpus_type,
file_extension='wav')
source_streams.append(curr_stream)
return source_streams
def __getitem__(self, index):
if self.sequence_mode:
# find the stream index and utterance index corresponding to the given index
stream_idx = -1
for i in range(len(self.source_stream_cum_sizes)):
if index < self.source_stream_cum_sizes[i]:
stream_idx = i
break
if stream_idx == -1:
raise Exception('index larger than available number of sentences. ')
if stream_idx==0:
utt_idx = index
else:
utt_idx = index - self.source_stream_cum_sizes[stream_idx-1]
data = self.data_buffer.generate((stream_idx, utt_idx))[0]
# data = self.data_buffer.get((stream_idx, utt_idx))
else:
data = self.data_buffer.get()
if self.transform is not None:
data = self.transform.apply(data, stream_keys=self.stream_idx_for_transform)
return data
def sample_in_seconds(self):
return self.sample_len_seconds
def __len__(self):
if self.sequence_mode:
return np.sum(self.source_stream_sizes)
else:
return self.data_buffer.get_len()
def __repr__(self):
fmt_str = 'Dataset ' + self.__class__.__name__ + '\n'
fmt_str += ' Number of datapoints: {}\n'.format(self.__len__())
tmp = 'train' if self.train is True else 'test'
fmt_str += ' Split: {}\n'.format(tmp)
fmt_str += ' Root Location: {}\n'.format(self.root)
tmp = ' Transforms (if any): '
fmt_str += '{0}{1}\n'.format(tmp, self.transform.__repr__().replace('\n', '\n' + ' ' * len(tmp)))
tmp = ' Target Transforms (if any): '
fmt_str += '{0}{1}'.format(tmp, self.target_transform.__repr__().replace('\n', '\n' + ' ' * len(tmp)))
return fmt_str
class DataGeneratorSequenceConfig:
"""
Define the configurations of data generation.
"""
def __init__(self, use_reverb, use_noise, snr_range, n_hour_per_epoch=10, sequence_mode=False, load_label=True, min_seglen=0, seglen=500, segshift=500, use_cmn=False, gain_norm=False, simulation_prob=0.5):
self.n_hour_per_epoch = n_hour_per_epoch
self.load_label = load_label
self.segment_config = {}
self.use_reverb = use_reverb
self.use_noise = use_noise
self.snr_range = snr_range
self.segment_config['sequence_mode'] = sequence_mode
self.segment_config['seglen'] = seglen # length of segments in terms of frames
self.segment_config['segshift'] = segshift
self.segment_config['min_seglen'] = min_seglen
self.n_segment_per_epoch = int(3600 * n_hour_per_epoch / self.segment_config['seglen'] * 100)
self.gain_norm = gain_norm
self.use_cmn = use_cmn
self.simulation_prob = simulation_prob
class DataGeneratorTrain:
"""
Generate simulated speech utterances from clean speech streams, noise streams, and rir streams. Responsible for
sampling of the data from the streams, and call SimpleSimulator to do the simulation. Also responsible for extract
features and make training samples.
"""
_window_file = 'mel80_window.txt' # the file that stores the Mel scale window coefficients
def __init__(self, source_streams, noise_streams, rir_streams, config, DEBUG=False):
"""
:param source_streams: a list of SpeechDataStream objects, containing the clean speech source files names and
meta-data such as label, utterance ID, and speaker ID.
:param noise_streams: a list of DataStream objects, containing noise file names.
:param rir_streams: a list of RIRDataStream objects, containing RIR file names and meta data information.
:param config: an object of type DataGeneratorSequenceConfig
:param DEBUG: if set to DEBUG mode, will plot the filterbanks and label.
"""
self._source_streams = source_streams
self._source_streams_prior = self._get_streams_prior(source_streams)
self._rir_streams = rir_streams
self._rir_streams_prior = self._get_streams_prior(rir_streams)
self._noise_streams = noise_streams
self._noise_streams_prior = self._get_streams_prior(noise_streams)
self._data_len = config.n_segment_per_epoch
self._single_source_simulator = SimpleSimulator(use_rir=config.use_reverb, use_noise=config.use_noise, snr_range=config.snr_range)
self._config = config
self._DEBUG = DEBUG
self._gen_window()
def _get_streams_prior(self, streams):
if streams is None:
return None
else:
n_entrys = np.asarray([stream.get_number_of_data() for stream in streams])
return n_entrys / np.sum(n_entrys)
def _gen_window(self):
# load the pre-computed window coefficients for 80D log filterbanks used in typical acoustic modeling.
# the window is computed by the following code
# import librosa
# self._window = librosa.filters.mel(16000, 512, n_mels=80, fmax=7690, htk=True)
mel_file = os.path.join(os.path.dirname(__file__), self._window_file)
with open(mel_file) as file:
lines = [line.rstrip('\n') for line in file]
self._window = np.vstack([np.asarray([np.float32(j) for j in i.split(",")]) for i in lines])
def _logfbank_extractor(self, wav):
# typical log fbank extraction for 16kHz speech data
preemphasis = 0.96
t1 = np.sum(self._window, 0)
t1[t1 == 0] = -1
inv = np.diag(1 / t1)
mel = self._window.dot(inv).T
wav = wav[1:] - preemphasis * wav[:-1]
S = stft(wav, n_fft=512, hop_length=160, win_length=400, window=np.hamming(400), center=False).T
spec_mag = np.abs(S)
spec_power = spec_mag ** 2
fbank_power = spec_power.T.dot(mel * 32768 ** 2) + 1
log_fbank = np.log(fbank_power)
return log_fbank
def generate(self, index=None):
"""
:param index: a tuple of 2 entries (source_stream_idx, utt_idx) that specifies which clean source file to use
for simulation. If not provided, will randomly choose one clean source file from the clean source streams.
:return: a list of training samples
"""
seg_len = self._config.segment_config['seglen']
seg_shift = self._config.segment_config['segshift']
if index is None: # if no index is given, let the simulator do random sampling
# sample a clean speech stream
source_stream_idx = np.random.choice(np.arange(len(self._source_streams)), replace=True, p=self._source_streams_prior)
# sample a clean speech utterance
_, utt_id, source_wav, _ = self._source_streams[source_stream_idx].sample_spk_and_utt(n_spk=1,
n_utt_per_spk=1,
load_data=True)
else: # if index is given, use the specified sentence
assert len(index) == 2
source_stream_idx = index[0]
utt_id = [self._source_streams[source_stream_idx].utt_id[index[1]]]
_, _, source_wav, _ = self._source_streams[source_stream_idx].read_utt_with_id(utt_id, load_data=True)
if np.random.random() > self._config.simulation_prob:
simulated_wav = source_wav[0]
else:
if self._noise_streams is None:
noise_wavs = None
else:
noise_stream_idx = np.random.choice(np.arange(len(self._noise_streams)), replace=True,
p=self._noise_streams_prior)
noise_wavs, noise_files = self._noise_streams[noise_stream_idx].sample_data()
if self._rir_streams is None:
source_rir = None
noise_rirs = None
else:
rir_stream_idx = np.random.choice(np.arange(len(self._rir_streams)), replace=True, p=self._rir_streams_prior)
n_rir = 1 if noise_wavs is None else 1+len(noise_wavs)
rir_wav, room_size, array_position, positions, t60 = self._rir_streams[rir_stream_idx].sample_rir(n_rir)
source_rir = rir_wav[0]
noise_rirs = rir_wav[1:]
simulated_wav, _, mask, config = self._single_source_simulator(source_wav[0],
dir_noise_wavs=noise_wavs,
source_rir=source_rir,
dir_noise_rirs=noise_rirs,
gen_mask=False, normalize_gain=self._config.gain_norm)
fbank = self._logfbank_extractor(simulated_wav[:,0])
if self._config.load_label:
_, label = self._source_streams[source_stream_idx].read_label_with_id(utt_id)
frame_label = label['label'][0].T
if 'aux_label' in label:
aux_label = label['aux_label']
else:
aux_label = np.zeros((1,1))
if np.abs(frame_label.shape[0] - fbank.shape[0])>5:
print("DataGeneratorTrain::generate: Warning: filterbank and label have significantly different number of frames. ")
n_fr = np.minimum(frame_label.shape[0], fbank.shape[0])
frame_label = frame_label[:n_fr,:]
fbank = fbank[:n_fr,:]
if self._config.use_cmn:
fbank = reader.preprocess.cmn(fbank, axis=0)
if self._config.segment_config['sequence_mode']:
if self._config.load_label:
train_samples = [(fbank, utt_id, frame_label, aux_label)]
else:
train_samples = [(fbank, utt_id)]
else:
fbank_seg = _utt2seg(fbank.T, seg_len, seg_shift)
if len(fbank_seg) == 0:
return []
if self._config.load_label:
label_seg = _utt2seg(frame_label.T, seg_len, seg_shift)
train_samples = [(fbank_seg[i].T, utt_id, label_seg[i].T) for i in range(len(label_seg))]
else:
train_samples = [(fbank_seg[i].T, utt_id) for i in range(len(fbank_seg))]
if self._DEBUG:
import matplotlib.pyplot as plt
n_sample = len(train_samples)
for i in range(n_sample):
plt.subplot(n_sample,2,i*2+1)
imagesc(train_samples[i][0].T)
plt.subplot(n_sample,2,i*2+2)
plt.plot(train_samples[i][2])
return train_samples
def get_len(self):
return self._data_len
