from Modules.vits.models import VitsModel, VitsTokenizer import sys import tempfile import re import os from collections import OrderedDict from Modules.hifigan import Decoder from Utils.PLBERT.util import load_plbert import phonemizer import torch from cached_path import cached_path import nltk import audresample nltk.download('punkt', download_dir='./') # comment if downloaded once nltk.download('punkt_tab', download_dir='./') nltk.data.path.append('.') import numpy as np import yaml import librosa from models import ProsodyPredictor, TextEncoder, StyleEncoder, MelSpec from nltk.tokenize import word_tokenize from Utils.text_utils import transliterate_number import textwrap device = 'cpu' if torch.cuda.is_available(): device = 'cuda' _pad = "$" _punctuation = ';:,.!?¡¿—…"«»“” ' _letters = 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz' _letters_ipa = "ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ" # Export all symbols: symbols = [_pad] + list(_punctuation) + list(_letters) + list(_letters_ipa) dicts = {} for i in range(len((symbols))): dicts[symbols[i]] = i class TextCleaner: def __init__(self, dummy=None): self.word_index_dictionary = dicts print(len(dicts)) def __call__(self, text): indexes = [] for char in text: try: indexes.append(self.word_index_dictionary[char]) except KeyError: print('CLEAN', text) return indexes textclenaer = TextCleaner() def alpha_num(f): f = re.sub(' +', ' ', f) # delete spaces f = re.sub(r'[^A-Z a-z0-9 ]+', '', f) # del non alpha num return f mel_spec = MelSpec().to(device) def compute_style(path): x, sr = librosa.load(path, sr=24000) x, _ = librosa.effects.trim(x, top_db=30) if sr != 24000: x = librosa.resample(x, sr, 24000) with torch.no_grad(): x = torch.from_numpy(x[None, :]).to(device=device, dtype=torch.float) mel_tensor = (torch.log(1e-5 + mel_spec(x)) + 4) / 4 #mel_tensor = preprocess(audio).to(device) ref_s = style_encoder(mel_tensor) ref_p = predictor_encoder(mel_tensor) # [bs, 11, 1, 128] s = torch.cat([ref_s, ref_p], dim=3) # [bs, 11, 1, 256] s = s[:, :, 0, :].transpose(1, 2) # [1, 128, 11] return s # [1, 128, 11] global_phonemizer = phonemizer.backend.EspeakBackend( language='en-us', preserve_punctuation=True, with_stress=True) # phonemizer = Phonemizer.from_checkpoint(str(cached_path('https://public-asai-dl-models.s3.eu-central-1.amazonaws.com/DeepPhonemizer/en_us_cmudict_ipa_forward.pt'))) args = yaml.safe_load(open(str('Utils/config.yml'))) ASR_config = args['ASR_config'] bert = load_plbert(args['PLBERT_dir']).eval().to(device) decoder = Decoder(dim_in=512, style_dim=128, dim_out=80, # n_mels resblock_kernel_sizes=[3, 7, 11], upsample_rates=[10, 5, 3, 2], upsample_initial_channel=512, resblock_dilation_sizes=[[1, 3, 5], [1, 3, 5], [1, 3, 5]], upsample_kernel_sizes=[20, 10, 6, 4]).eval().to(device) text_encoder = TextEncoder(channels=512, kernel_size=5, depth=3, # args['model_params']['n_layer'], n_symbols=178, # args['model_params']['n_token'] ).eval().to(device) predictor = ProsodyPredictor(style_dim=128, d_hid=512, nlayers=3, # OFFICIAL config.nlayers=5; max_dur=50).eval().to(device) style_encoder = StyleEncoder(dim_in=64, style_dim=128, max_conv_dim=512).eval().to(device) # acoustic style encoder predictor_encoder = StyleEncoder(dim_in=64, style_dim=128, max_conv_dim=512).eval().to(device) # prosodic style encoder bert_encoder = torch.nn.Linear(bert.config.hidden_size, 512).eval().to(device) # params_whole = torch.load('freevc2/yl4579_styletts2.pth' map_location='cpu') params_whole = torch.load(str(cached_path( "hf://yl4579/StyleTTS2-LibriTTS/Models/LibriTTS/epochs_2nd_00020.pth")), map_location='cpu', weights_only=True) params = params_whole['net'] #params['decoder'].pop('module.generator.m_source.l_linear.weight') #params['decoder'].pop('module.generator.m_source.l_linear.bias') # SourceHNSf def _del_prefix(d): # del ".module" out = OrderedDict() for k, v in d.items(): out[k[7:]] = v return out bert.load_state_dict(_del_prefix(params['bert']), strict=True) bert_encoder.load_state_dict(_del_prefix(params['bert_encoder']), strict=True) # XTRA non-ckpt LSTMs nlayers add slowiness to voice predictor.load_state_dict(_del_prefix(params['predictor']), strict=True) decoder.load_state_dict(_del_prefix(params['decoder']), strict=True) text_encoder.load_state_dict(_del_prefix(params['text_encoder']), strict=True) predictor_encoder.load_state_dict(_del_prefix( params['predictor_encoder']), strict=True) style_encoder.load_state_dict(_del_prefix( params['style_encoder']), strict=True) def inference(text, ref_s): # text = transliterate_number(text, lang='en').strip() # Transliteration only used for foreign() # perhaps add xtra . after ? ; ps = global_phonemizer.phonemize([text]) ps = word_tokenize(ps[0]) ps = ' '.join(ps) tokens = textclenaer(ps) tokens.insert(0, 0) tokens = torch.LongTensor(tokens).to(device).unsqueeze(0) with torch.no_grad(): hidden_states = text_encoder(tokens) bert_dur = bert(tokens, attention_mask=torch.ones_like(tokens)) d_en = bert_encoder(bert_dur).transpose(-1, -2) aln_trg, F0_pred, N_pred = predictor(d_en=d_en, s=ref_s[:, 128:, :]) asr = torch.bmm(aln_trg, hidden_states) asr = asr.transpose(1, 2) asr = torch.cat([asr[:, :, 0:1], asr[:, :, 0:-1]], 2) x = decoder(asr=asr, # [1, 512, 201] F0_curve=F0_pred, # [1, 1, 402] 2x time N=N_pred, # [1, 1, 402] 2x time s=ref_s[:, :128, :]) # [1, 256, 1] x = x.cpu().numpy()[0, 0, :] x[-400:] = 0 # noisy pulse produced for unterminated sentences, in absence of punctuation, (not sure if same behaviour for all voices) # StyleTTS2 is 24kHz -> Resample to 16kHz as is AudioGen / MMS if x.shape[0] > 10: x = audresample.resample(signal=x.astype(np.float32), original_rate=24000, target_rate=16000)[0, :] # audresample reshapes (64,) -> (1,64) | Volume Normalisation applies in api.py:tts_multi_sentence() else: print('\n\n\n\n\nEMPTY TTS\n\n\n\n\n\nn', x.shape) x = np.zeros(0) return x # ___________________________________________________________ # https://huggingface.co/spaces/mms-meta/MMS/blob/main/tts.py # ___________________________________________________________ # -*- coding: utf-8 -*- # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. TTS_LANGUAGES = {} # with open('_d.csv', 'w') as f2: with open(f"Utils/all_langs.csv") as f: for line in f: iso, name = line.split(",", 1) TTS_LANGUAGES[iso.strip()] = name.strip() # f2.write(iso + ',' + name.replace("a S","")+'\n') # ============================================================================================== # LOAD hun / ron / serbian - rmc-script_latin / cyrillic-Carpathian (not Vlax) # ============================================================================================== PHONEME_MAP = { 'služ': 'sloooozz', # 'službeno' 'suver': 'siuveeerra', # 'suverena' 'država': 'dirrezav', # 'država' 'iči': 'ici', # 'Graniči' 's ': 'se', # a s with space 'q': 'ku', 'w': 'aou', 'z': 's', "š": "s", 'th': 'ta', 'v': 'vv', # "ć": "č", # "đ": "ď", # "lj": "ľ", # "nj": "ň", "ž": "z", # "c": "č" } def fix_phones(text): for src, target in PHONEME_MAP.items(): text = text.replace(src, target) # text = re.sub(r'\s+', '` `', text) #.strip() #.lower() # text = re.sub(r'\s+', '_ _', text) # almost proper pausing return text.replace(',', '_ _').replace('.', '_ _') def has_cyrillic(text): # https://stackoverflow.com/questions/48255244/python-check-if-a-string-contains-cyrillic-characters return bool(re.search('[\u0400-\u04FF]', text)) def foreign(text=None, # split sentences here so we can prepend a txt for german to each sentence to # fall on the male voice (Sink attn) lang='romanian', speed=None): # https://huggingface.co/dkounadis/artificial-styletts2/blob/main/Utils/all_langs.csv lang = lang.lower() # https://huggingface.co/spaces/mms-meta/MMS if 'hun' in lang: lang_code = 'hun' elif any([i in lang for i in ['ser', 'bosn', 'herzegov', 'montenegr', 'macedon']]): if has_cyrillic(text): # check 0-th sentence if is cyrillic # romani carpathian (also has latin / cyrillic Vlax) lang_code = 'rmc-script_cyrillic' else: # romani carpathian (has also Vlax) lang_code = 'rmc-script_latin' elif 'rom' in lang: lang_code = 'ron' elif 'ger' in lang or 'deu' in lang or 'allem' in lang: lang_code = 'deu' elif 'alban' in lang: lang_code = 'sqi' else: lang_code = lang.split()[0].strip() # load VITS # net_g = VitsModel.from_pretrained(f'facebook/mms-tts-{lang_code}').eval().to(device) # tokenizer = VitsTokenizer.from_pretrained(f'facebook/mms-tts-{lang_code}') global cached_lang_code, cached_net_g, cached_tokenizer if 'cached_lang_code' not in globals() or cached_lang_code != lang_code: cached_lang_code = lang_code cached_net_g = VitsModel.from_pretrained(f'facebook/mms-tts-{lang_code}').eval().to(device) cached_tokenizer = VitsTokenizer.from_pretrained(f'facebook/mms-tts-{lang_code}') net_g = cached_net_g tokenizer = cached_tokenizer total_audio = [] # Split long sentences if deu to control voice switch - for other languages let text no-split if not isinstance(text, list): # Split Very long sentences text = [sub_sent+' ' for sub_sent in textwrap.wrap(text, 440, break_long_words=0)] for _t in text: _t = _t.lower() # NUMBERS try: _t = transliterate_number(_t, lang=lang_code) except NotImplementedError: print('Transliterate Numbers - NotImplemented for {lang_code=}', _t,'\n____________________________________________') # PRONOUNC. if lang_code == 'rmc-script_latin': _t = fix_phones(_t) # phonemes replace per language elif lang_code == 'ron': # tone _t = _t.replace("ţ", "ț" ).replace('ț', 'ts').replace('î', 'u').replace('â', 'a').replace('ş', 's') # /data/dkounadis/.hf7/hub/models--facebook--mms-tts/snapshots/44cc7fb408064ef9ea6e7c59130d88cac1274671/models/rmc-script_latin/vocab.txt # input_ids / attention_mask inputs = tokenizer(_t, return_tensors="pt") with torch.no_grad(): # MMS x = net_g(input_ids=inputs.input_ids.to(device), attention_mask=inputs.attention_mask.to(device), lang_code=lang_code, )[0, :] # crop the 1st audio - is PREFIX text 156000 samples to chose deu voice / VitsAttention() total_audio.append(x) print(f'\n\n_______________________________ {_t} {x.shape=}') x = torch.cat(total_audio).cpu().numpy() # x /= np.abs(x).max() + 1e-7 ~ Volume normalisation @api.py:tts_multi_sentence() OR demo.py return x # 16kHz - only resample StyleTTS2 from 24Hkz -> 16kHz