import os import os import random import sys import torch import torch.nn.functional as F import torch.utils.data import torchaudio from tqdm import tqdm from data.audio.unsupervised_audio_dataset import load_audio, load_similar_clips from models.tacotron2.taco_utils import load_filepaths_and_text from models.tacotron2.text import text_to_sequence, sequence_to_text from utils.util import opt_get def load_tsv(filename): with open(filename, encoding='utf-8') as f: components = [line.strip().split('\t') for line in f] base = os.path.dirname(filename) filepaths_and_text = [[os.path.join(base, f'{component[1]}'), component[0]] for component in components] return filepaths_and_text def load_mozilla_cv(filename): with open(filename, encoding='utf-8') as f: components = [line.strip().split('\t') for line in f][1:] # First line is the header base = os.path.dirname(filename) filepaths_and_text = [[os.path.join(base, f'clips/{component[1]}'), component[2]] for component in components] return filepaths_and_text def load_voxpopuli(filename): with open(filename, encoding='utf-8') as f: lines = [line.strip().split('\t') for line in f][1:] # First line is the header base = os.path.dirname(filename) filepaths_and_text = [] for line in lines: if len(line) == 0: continue file, raw_text, norm_text, speaker_id, split, gender = line year = file[:4] filepaths_and_text.append([os.path.join(base, year, f'{file}.ogg.wav'), raw_text]) return filepaths_and_text class CharacterTokenizer: def encode(self, txt): return text_to_sequence(txt, ['english_cleaners']) def decode(self, seq): return sequence_to_text(seq) class TextWavLoader(torch.utils.data.Dataset): def __init__(self, hparams): self.path = hparams['path'] if not isinstance(self.path, list): self.path = [self.path] fetcher_mode = opt_get(hparams, ['fetcher_mode'], 'lj') if not isinstance(fetcher_mode, list): fetcher_mode = [fetcher_mode] assert len(self.path) == len(fetcher_mode) self.load_conditioning = opt_get(hparams, ['load_conditioning'], False) self.conditioning_candidates = opt_get(hparams, ['num_conditioning_candidates'], 1) self.conditioning_length = opt_get(hparams, ['conditioning_length'], 44100) self.debug_failures = opt_get(hparams, ['debug_loading_failures'], False) self.audiopaths_and_text = [] for p, fm in zip(self.path, fetcher_mode): if fm == 'lj' or fm == 'libritts': fetcher_fn = load_filepaths_and_text elif fm == 'tsv': fetcher_fn = load_tsv elif fm == 'mozilla_cv': assert not self.load_conditioning # Conditioning inputs are incompatible with mozilla_cv fetcher_fn = load_mozilla_cv elif fm == 'voxpopuli': assert not self.load_conditioning # Conditioning inputs are incompatible with voxpopuli fetcher_fn = load_voxpopuli else: raise NotImplementedError() self.audiopaths_and_text.extend(fetcher_fn(p)) self.text_cleaners = hparams.text_cleaners self.sample_rate = hparams.sample_rate random.seed(hparams.seed) random.shuffle(self.audiopaths_and_text) self.max_wav_len = opt_get(hparams, ['max_wav_length'], None) self.max_text_len = opt_get(hparams, ['max_text_length'], None) assert self.max_wav_len is not None and self.max_text_len is not None self.use_bpe_tokenizer = opt_get(hparams, ['use_bpe_tokenizer'], True) if self.use_bpe_tokenizer: from data.audio.voice_tokenizer import VoiceBpeTokenizer self.tokenizer = VoiceBpeTokenizer(opt_get(hparams, ['tokenizer_vocab'], '../experiments/bpe_lowercase_asr_256.json')) else: self.tokenizer = CharacterTokenizer() self.skipped_items = 0 # records how many items are skipped when accessing an index. def get_wav_text_pair(self, audiopath_and_text): # separate filename and text audiopath, text = audiopath_and_text[0], audiopath_and_text[1] text_seq = self.get_text(text) wav = load_audio(audiopath, self.sample_rate) return (text_seq, wav, text, audiopath_and_text[0]) def get_text(self, text): tokens = self.tokenizer.encode(text) tokens = torch.IntTensor(tokens) if self.use_bpe_tokenizer: # Assert if any UNK,start tokens encountered. assert not torch.any(tokens == 1) # The stop token should always be sacred. assert not torch.any(tokens == 0) return tokens def __getitem__(self, index): self.skipped_items += 1 try: tseq, wav, text, path = self.get_wav_text_pair(self.audiopaths_and_text[index]) cond, cond_is_self = load_similar_clips(self.audiopaths_and_text[index][0], self.conditioning_length, self.sample_rate, n=self.conditioning_candidates) if self.load_conditioning else None except: if self.skipped_items > 100: raise # Rethrow if we have nested too far. if self.debug_failures: print(f"error loading {self.audiopaths_and_text[index][0]} {sys.exc_info()}") return self[(index+1) % len(self)] actually_skipped_items = self.skipped_items self.skipped_items = 0 if wav is None or \ (self.max_wav_len is not None and wav.shape[-1] > self.max_wav_len) or \ (self.max_text_len is not None and tseq.shape[0] > self.max_text_len): # Basically, this audio file is nonexistent or too long to be supported by the dataset. # It's hard to handle this situation properly. Best bet is to return the a random valid token and skew the dataset somewhat as a result. if self.debug_failures: print(f"error loading {path}: ranges are out of bounds; {wav.shape[-1]}, {tseq.shape[0]}") rv = random.randint(0,len(self)-1) return self[rv] orig_output = wav.shape[-1] orig_text_len = tseq.shape[0] if wav.shape[-1] != self.max_wav_len: wav = F.pad(wav, (0, self.max_wav_len - wav.shape[-1])) if tseq.shape[0] != self.max_text_len: tseq = F.pad(tseq, (0, self.max_text_len - tseq.shape[0])) res = { 'real_text': text, 'padded_text': tseq, 'text_lengths': torch.tensor(orig_text_len, dtype=torch.long), 'wav': wav, 'wav_lengths': torch.tensor(orig_output, dtype=torch.long), 'filenames': path, 'skipped_items': actually_skipped_items, } if self.load_conditioning: res['conditioning'] = cond res['conditioning_contains_self'] = cond_is_self return res def __len__(self): return len(self.audiopaths_and_text) if __name__ == '__main__': batch_sz = 8 params = { 'mode': 'paired_voice_audio', 'path': ['Y:\\bigasr_dataset\\hifi_tts\\test.txt'], 'fetcher_mode': ['libritts'], 'phase': 'train', 'n_workers': 0, 'batch_size': batch_sz, 'max_wav_length': 255995, 'max_text_length': 200, 'sample_rate': 22050, 'load_conditioning': True, 'num_conditioning_candidates': 2, 'conditioning_length': 44000, 'use_bpe_tokenizer': True, } from data import create_dataset, create_dataloader def save(b, i, ib, key, c=None): if c is not None: torchaudio.save(f'{i}_clip_{ib}_{key}_{c}.wav', b[key][ib][c], 22050) else: torchaudio.save(f'{i}_clip_{ib}_{key}.wav', b[key][ib], 22050) ds, c = create_dataset(params, return_collate=True) dl = create_dataloader(ds, params, collate_fn=c) i = 0 m = None for i, b in tqdm(enumerate(dl)): for ib in range(batch_sz): print(f'{i} {ib} {b["real_text"][ib]}') save(b, i, ib, 'wav') if i > 5: break