File size: 8,997 Bytes
0a948c1 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 |
# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from functools import partial
import onnxruntime
import torch
import numpy as np
import whisper
from typing import Callable
import torchaudio.compliance.kaldi as kaldi
import torchaudio
import os
import re
import inflect
try:
import ttsfrd
use_ttsfrd = True
except ImportError:
print("failed to import ttsfrd, use WeTextProcessing instead")
from tn.chinese.normalizer import Normalizer as ZhNormalizer
from tn.english.normalizer import Normalizer as EnNormalizer
use_ttsfrd = False
from cosyvoice.utils.frontend_utils import contains_chinese, replace_blank, replace_corner_mark, remove_bracket, spell_out_number, split_paragraph
class CosyVoiceFrontEnd:
def __init__(self,
get_tokenizer: Callable,
feat_extractor: Callable,
campplus_model: str,
speech_tokenizer_model: str,
spk2info: str = '',
instruct: bool = False,
allowed_special: str = 'all'):
self.tokenizer = get_tokenizer()
self.feat_extractor = feat_extractor
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
option = onnxruntime.SessionOptions()
option.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
option.intra_op_num_threads = 1
self.campplus_session = onnxruntime.InferenceSession(campplus_model, sess_options=option, providers=["CPUExecutionProvider"])
self.speech_tokenizer_session = onnxruntime.InferenceSession(speech_tokenizer_model, sess_options=option, providers=["CUDAExecutionProvider"if torch.cuda.is_available() else "CPUExecutionProvider"])
if os.path.exists(spk2info):
self.spk2info = torch.load(spk2info, map_location=self.device)
self.instruct = instruct
self.allowed_special = allowed_special
self.inflect_parser = inflect.engine()
self.use_ttsfrd = use_ttsfrd
if self.use_ttsfrd:
self.frd = ttsfrd.TtsFrontendEngine()
ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
assert self.frd.initialize('{}/../../pretrained_models/CosyVoice-ttsfrd/resource'.format(ROOT_DIR)) is True, 'failed to initialize ttsfrd resource'
self.frd.set_lang_type('pinyin')
self.frd.enable_pinyin_mix(True)
self.frd.set_breakmodel_index(1)
else:
self.zh_tn_model = ZhNormalizer(remove_erhua=False, full_to_half=False)
self.en_tn_model = EnNormalizer()
def _extract_text_token(self, text):
text_token = self.tokenizer.encode(text, allowed_special=self.allowed_special)
text_token = torch.tensor([text_token], dtype=torch.int32).to(self.device)
text_token_len = torch.tensor([text_token.shape[1]], dtype=torch.int32).to(self.device)
return text_token, text_token_len
def _extract_speech_token(self, speech):
feat = whisper.log_mel_spectrogram(speech, n_mels=128)
speech_token = self.speech_tokenizer_session.run(None, {self.speech_tokenizer_session.get_inputs()[0].name: feat.detach().cpu().numpy(),
self.speech_tokenizer_session.get_inputs()[1].name: np.array([feat.shape[2]], dtype=np.int32)})[0].flatten().tolist()
speech_token = torch.tensor([speech_token], dtype=torch.int32).to(self.device)
speech_token_len = torch.tensor([speech_token.shape[1]], dtype=torch.int32).to(self.device)
return speech_token, speech_token_len
def _extract_spk_embedding(self, speech):
feat = kaldi.fbank(speech,
num_mel_bins=80,
dither=0,
sample_frequency=16000)
feat = feat - feat.mean(dim=0, keepdim=True)
embedding = self.campplus_session.run(None, {self.campplus_session.get_inputs()[0].name: feat.unsqueeze(dim=0).cpu().numpy()})[0].flatten().tolist()
embedding = torch.tensor([embedding]).to(self.device)
return embedding
def _extract_speech_feat(self, speech):
speech_feat = self.feat_extractor(speech).squeeze(dim=0).transpose(0, 1).to(self.device)
speech_feat = speech_feat.unsqueeze(dim=0)
speech_feat_len = torch.tensor([speech_feat.shape[1]], dtype=torch.int32).to(self.device)
return speech_feat, speech_feat_len
def text_normalize(self, text, split=True):
text = text.strip()
if contains_chinese(text):
if self.use_ttsfrd:
text = self.frd.get_frd_extra_info(text, 'input')
else:
text = self.zh_tn_model.normalize(text)
text = text.replace("\n", "")
text = replace_blank(text)
text = replace_corner_mark(text)
text = text.replace(".", "、")
text = text.replace(" - ", "，")
text = remove_bracket(text)
text = re.sub(r'[，,]+$', '。', text)
texts = [i for i in split_paragraph(text, partial(self.tokenizer.encode, allowed_special=self.allowed_special), "zh", token_max_n=80,
token_min_n=60, merge_len=20,
comma_split=False)]
else:
if self.use_ttsfrd:
text = self.frd.get_frd_extra_info(text, 'input')
else:
text = self.en_tn_model.normalize(text)
text = spell_out_number(text, self.inflect_parser)
texts = [i for i in split_paragraph(text, partial(self.tokenizer.encode, allowed_special=self.allowed_special), "en", token_max_n=80,
token_min_n=60, merge_len=20,
comma_split=False)]
if split is False:
return text
return texts
def frontend_sft(self, tts_text, spk_id):
tts_text_token, tts_text_token_len = self._extract_text_token(tts_text)
embedding = self.spk2info[spk_id]['embedding']
model_input = {'text': tts_text_token, 'text_len': tts_text_token_len, 'llm_embedding': embedding, 'flow_embedding': embedding}
return model_input
def frontend_zero_shot(self, tts_text, prompt_text, prompt_speech_16k):
tts_text_token, tts_text_token_len = self._extract_text_token(tts_text)
prompt_text_token, prompt_text_token_len = self._extract_text_token(prompt_text)
prompt_speech_22050 = torchaudio.transforms.Resample(orig_freq=16000, new_freq=22050)(prompt_speech_16k)
speech_feat, speech_feat_len = self._extract_speech_feat(prompt_speech_22050)
speech_token, speech_token_len = self._extract_speech_token(prompt_speech_16k)
embedding = self._extract_spk_embedding(prompt_speech_16k)
model_input = {'text': tts_text_token, 'text_len': tts_text_token_len,
'prompt_text': prompt_text_token, 'prompt_text_len': prompt_text_token_len,
'llm_prompt_speech_token': speech_token, 'llm_prompt_speech_token_len': speech_token_len,
'flow_prompt_speech_token': speech_token, 'flow_prompt_speech_token_len': speech_token_len,
'prompt_speech_feat': speech_feat, 'prompt_speech_feat_len': speech_feat_len,
'llm_embedding': embedding, 'flow_embedding': embedding}
return model_input
def frontend_cross_lingual(self, tts_text, prompt_speech_16k):
model_input = self.frontend_zero_shot(tts_text, '', prompt_speech_16k)
# in cross lingual mode, we remove prompt in llm
del model_input['prompt_text']
del model_input['prompt_text_len']
del model_input['llm_prompt_speech_token']
del model_input['llm_prompt_speech_token_len']
return model_input
def frontend_instruct(self, tts_text, spk_id, instruct_text):
model_input = self.frontend_sft(tts_text, spk_id)
# in instruct mode, we remove spk_embedding in llm due to information leakage
del model_input['llm_embedding']
instruct_text_token, instruct_text_token_len = self._extract_text_token(instruct_text + '<endofprompt>')
model_input['prompt_text'] = instruct_text_token
model_input['prompt_text_len'] = instruct_text_token_len
return model_input
|