def __init__(self, working_dir: str, max_length: int, model: DeepSpeakerModel):
self.working_dir = working_dir
self.audio = Audio(cache_dir=working_dir)
logger.info(f'Picking audio from {working_dir}.')
self.sp_to_utt_train = train_test_sp_to_utt(self.audio, is_test=False)
self.sp_to_utt_test = train_test_sp_to_utt(self.audio, is_test=True)
self.max_length = max_length
self.model = model
self.nb_per_speaker = 2
self.nb_speakers = 640
self.history_length = 4
self.history_every = 100 # batches.
self.total_history_length = self.nb_speakers * self.nb_per_speaker * self.history_length # 25,600
self.metadata_train_speakers = Counter()
self.metadata_output_file = os.path.join(self.working_dir, 'debug_batcher.json')
self.history_embeddings_train = deque(maxlen=self.total_history_length)
self.history_utterances_train = deque(maxlen=self.total_history_length)
self.history_model_inputs_train = deque(maxlen=self.total_history_length)
self.history_embeddings = None
self.history_utterances = None
self.history_model_inputs = None
self.batch_count = 0
for _ in tqdm(range(self.history_length), desc='Initializing the batcher'): # init history.
self.update_triplets_history()
def generate_per_phase(self,
max_length=NUM_FRAMES,
num_per_speaker=3000,
is_test=False):
# train OR test.
num_speakers = len(self.audio.speaker_ids)
sp_to_utt = train_test_sp_to_utt(self.audio, is_test)
# 64 fbanks 1 channel(s).
# float32
kx = np.zeros(
(num_speakers * num_per_speaker, max_length, NUM_FBANKS, 1),
dtype=np.float32)
ky = np.zeros((num_speakers * num_per_speaker, 1), dtype=np.float32)
kg = np.zeros((num_speakers * num_per_speaker, 1), dtype=np.float32)
desc = f'Converting to Keras format [{"test" if is_test else "train"}]'
for i, speaker_id in enumerate(tqdm(self.audio.speaker_ids,
desc=desc)):
utterances_files = sp_to_utt[speaker_id]
for j, utterance_file in enumerate(
np.random.choice(utterances_files,
size=num_per_speaker,
replace=True)):
self.load_into_mat(utterance_file, self.categorical_speakers,
speaker_id, max_length, kx, ky, kg,
i * num_per_speaker + j)
return kx, ky, kg
chunk_size in ms
iterate over chunks until you find the first one with sound
'''
trim_ms = 0 # ms
assert chunk_size > 0 # to avoid infinite loop
while sound[trim_ms:trim_ms + chunk_size].dBFS < silence_threshold and trim_ms < len(sound):
trim_ms += chunk_size
return trim_ms
deep_speaker_root = os.getenv('WORKING_DIR') or os.path.join(os.getenv('HOME'), '.deep-speaker-wd')
audio = Audio(cache_dir=os.path.join(deep_speaker_root, 'triplet-training'))
sp_to_utt_test = train_test_sp_to_utt(audio, is_test=True)
sp_to_utt_train = train_test_sp_to_utt(audio, is_test=False)
for speaker in audio.speaker_ids:
sp_to_utt_test[speaker] += sp_to_utt_train[speaker][5:]
print("No. audio: %d" % sum(len(sp_to_utt_test[s]) for s in sp_to_utt_test))
root = os.path.join('samples', 'librispeech')
os.makedirs(root, exist_ok=True)
outputs = []
def load_transcripts():
transcripts = dict()
for transcript_path in tqdm(glob.glob(os.path.join(deep_speaker_root, 'LibriSpeech/**/**/**/*.txt'))):
import glob
import os
import re
import subprocess
from pydub import AudioSegment
from tqdm import tqdm
from audio import Audio
from utils import train_test_sp_to_utt
deep_speaker_root = '/home/trungvd/.deep-speaker-wd'
audio = Audio(cache_dir=os.path.join(deep_speaker_root, 'triplet-training'))
sp_to_utt_test = train_test_sp_to_utt(audio, is_test=True)
root = '/home/trungvd/repos/speech-reconstruction/samples/librispeech'
os.makedirs(root, exist_ok=True)
outputs = []
def load_transcripts():
transcripts = dict()
for transcript_path in tqdm(
glob.glob(
'/home/trungvd/.deep-speaker-wd/LibriSpeech/**/**/**/*.txt')):
with open(transcript_path, 'r') as f:
lines = f.read().strip().split('\n')
for line in lines:
ids, trans = line.split(' ', 1)
transcripts[ids] = trans
return transcripts