deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-04 09:29:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

add some instruction and assert

Browse Source
This commit is contained in:
lyuxiang.lx
2024-12-30 16:41:57 +08:00
parent bfcbc73df8
commit b9ddcba5fd
6 changed files with 52 additions and 63 deletions
Download Patch File Download Diff File Expand all files Collapse all files

42
cosyvoice/cli/cosyvoice.py
View File

@@ -20,23 +20,24 @@ import torch
from cosyvoice.cli.frontend import CosyVoiceFrontEnd
from cosyvoice.cli.model import CosyVoiceModel, CosyVoice2Model
from cosyvoice.utils.file_utils import logging
from cosyvoice.utils.class_utils import get_model_type
class CosyVoice:
def __init__(self, model_dir, load_jit=True, load_onnx=False, fp16=True):
instruct = True if '-Instruct' in model_dir else False
self.instruct = True if '-Instruct' in model_dir else False
self.model_dir = model_dir
if not os.path.exists(model_dir):
model_dir = snapshot_download(model_dir)
with open('{}/cosyvoice.yaml'.format(model_dir), 'r') as f:
configs = load_hyperpyyaml(f)
assert get_model_type(configs) == CosyVoiceModel, 'do not use {} for CosyVoice initialization!'.format(model_dir)
self.frontend = CosyVoiceFrontEnd(configs['get_tokenizer'],
configs['feat_extractor'],
'{}/campplus.onnx'.format(model_dir),
'{}/speech_tokenizer_v1.onnx'.format(model_dir),
'{}/spk2info.pt'.format(model_dir),
instruct,
configs['allowed_special'])
self.sample_rate = configs['sample_rate']
if torch.cuda.is_available() is False and (fp16 is True or load_jit is True):
@@ -85,8 +86,6 @@ class CosyVoice:
start_time = time.time()
def inference_cross_lingual(self, tts_text, prompt_speech_16k, stream=False, speed=1.0, text_frontend=True):
if self.frontend.instruct is True and isinstance(self.model, CosyVoiceModel):
raise ValueError('{} do not support cross_lingual inference'.format(self.model_dir))
for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)):
model_input = self.frontend.frontend_cross_lingual(i, prompt_speech_16k, self.sample_rate)
start_time = time.time()
@@ -98,8 +97,8 @@ class CosyVoice:
start_time = time.time()
def inference_instruct(self, tts_text, spk_id, instruct_text, stream=False, speed=1.0, text_frontend=True):
assert isinstance(self.model, CosyVoiceModel)
if self.frontend.instruct is False:
assert isinstance(self.model, CosyVoiceModel), 'inference_instruct is only implemented for CosyVoice!'
if self.instruct is False:
raise ValueError('{} do not support instruct inference'.format(self.model_dir))
instruct_text = self.frontend.text_normalize(instruct_text, split=False, text_frontend=text_frontend)
for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)):
@@ -112,18 +111,6 @@ class CosyVoice:
yield model_output
start_time = time.time()
def inference_instruct2(self, tts_text, instruct_text, prompt_speech_16k, stream=False, speed=1.0, text_frontend=True):
assert isinstance(self.model, CosyVoice2Model)
for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)):
model_input = self.frontend.frontend_instruct2(i, instruct_text, prompt_speech_16k, self.sample_rate)
start_time = time.time()
logging.info('synthesis text {}'.format(i))
for model_output in self.model.tts(**model_input, stream=stream, speed=speed):
speech_len = model_output['tts_speech'].shape[1] / self.sample_rate
logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
yield model_output
start_time = time.time()
def inference_vc(self, source_speech_16k, prompt_speech_16k, stream=False, speed=1.0):
model_input = self.frontend.frontend_vc(source_speech_16k, prompt_speech_16k, self.sample_rate)
start_time = time.time()
@@ -137,18 +124,18 @@ class CosyVoice:
class CosyVoice2(CosyVoice):
def __init__(self, model_dir, load_jit=False, load_onnx=False, load_trt=False):
instruct = True if '-Instruct' in model_dir else False
self.instruct = True if '-Instruct' in model_dir else False
self.model_dir = model_dir
if not os.path.exists(model_dir):
model_dir = snapshot_download(model_dir)
with open('{}/cosyvoice.yaml'.format(model_dir), 'r') as f:
configs = load_hyperpyyaml(f, overrides={'qwen_pretrain_path': os.path.join(model_dir, 'CosyVoice-BlankEN')})
assert get_model_type(configs) == CosyVoice2Model, 'do not use {} for CosyVoice2 initialization!'.format(model_dir)
self.frontend = CosyVoiceFrontEnd(configs['get_tokenizer'],
configs['feat_extractor'],
'{}/campplus.onnx'.format(model_dir),
'{}/speech_tokenizer_v2.onnx'.format(model_dir),
'{}/spk2info.pt'.format(model_dir),
instruct,
configs['allowed_special'])
self.sample_rate = configs['sample_rate']
if torch.cuda.is_available() is False and load_jit is True:
@@ -168,3 +155,18 @@ class CosyVoice2(CosyVoice):
if load_trt:
self.model.load_trt('{}/flow.decoder.estimator.fp16.Volta.plan'.format(model_dir))
del configs
def inference_instruct(self, *args, **kwargs):
raise NotImplementedError('inference_instruct is not implemented for CosyVoice2!')
def inference_instruct2(self, tts_text, instruct_text, prompt_speech_16k, stream=False, speed=1.0, text_frontend=True):
assert isinstance(self.model, CosyVoice2Model), 'inference_instruct2 is only implemented for CosyVoice2!'
for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)):
model_input = self.frontend.frontend_instruct2(i, instruct_text, prompt_speech_16k, self.sample_rate)
start_time = time.time()
logging.info('synthesis text {}'.format(i))
for model_output in self.model.tts(**model_input, stream=stream, speed=speed):
speech_len = model_output['tts_speech'].shape[1] / self.sample_rate
logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
yield model_output
start_time = time.time()