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

update

Browse Source
This commit is contained in:
lyuxiang.lx
2024-12-12 16:46:28 +08:00
parent 2345ce6be2
commit c693039d14
6 changed files with 145 additions and 71 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
cosyvoice/cli/cosyvoice.py
View File

@@ -38,6 +38,7 @@ class CosyVoice:
'{}/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):
load_jit = False
fp16 = False
@@ -64,7 +65,7 @@ class CosyVoice:
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] / 22050
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()
@@ -74,11 +75,11 @@ class CosyVoice:
for i in tqdm(self.frontend.text_normalize(tts_text, split=True)):
if len(i) < 0.5 * len(prompt_text):
logging.warning('synthesis text {} too short than prompt text {}, this may lead to bad performance'.format(i, prompt_text))
model_input = self.frontend.frontend_zero_shot(i, prompt_text, prompt_speech_16k)
model_input = self.frontend.frontend_zero_shot(i, prompt_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] / 22050
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()
@@ -87,11 +88,11 @@ class CosyVoice:
if self.frontend.instruct is True:
raise ValueError('{} do not support cross_lingual inference'.format(self.model_dir))
for i in tqdm(self.frontend.text_normalize(tts_text, split=True)):
model_input = self.frontend.frontend_cross_lingual(i, prompt_speech_16k)
model_input = self.frontend.frontend_cross_lingual(i, 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] / 22050
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()
@@ -105,23 +106,23 @@ class CosyVoice:
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] / 22050
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)
model_input = self.frontend.frontend_vc(source_speech_16k, prompt_speech_16k, self.sample_rate)
start_time = time.time()
for model_output in self.model.vc(**model_input, stream=stream, speed=speed):
speech_len = model_output['tts_speech'].shape[1] / 22050
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()
class CosyVoice2(CosyVoice):
def __init__(self, model_dir, load_jit=True, load_onnx=False, fp16=True):
def __init__(self, model_dir, load_jit=False, load_onnx=False, load_trt=False):
instruct = True if '-Instruct' in model_dir else False
self.model_dir = model_dir
if not os.path.exists(model_dir):
@@ -135,18 +136,21 @@ class CosyVoice2(CosyVoice):
'{}/spk2info.pt'.format(model_dir),
instruct,
configs['allowed_special'])
if torch.cuda.is_available() is False and (fp16 is True or load_jit is True):
self.sample_rate = configs['sample_rate']
if torch.cuda.is_available() is False and load_jit is True:
load_jit = False
fp16 = False
logging.warning('cpu do not support fp16 and jit, force set to False')
self.model = CosyVoice2Model(configs['llm'], configs['flow'], configs['hift'], fp16)
logging.warning('cpu do not support jit, force set to False')
self.model = CosyVoice2Model(configs['llm'], configs['flow'], configs['hift'])
self.model.load('{}/llm.pt'.format(model_dir),
'{}/flow.pt'.format(model_dir),
'{}/hift.pt'.format(model_dir))
if load_jit:
self.model.load_jit('{}/llm.text_encoder.fp16.zip'.format(model_dir),
'{}/llm.llm.fp16.zip'.format(model_dir),
'{}/flow.encoder.fp32.zip'.format(model_dir))
self.model.load_jit('{}/flow.encoder.fp32.zip'.format(model_dir))
if load_trt is True and load_onnx is True:
load_onnx = False
logging.warning('can not set both load_trt and load_onnx to True, force set load_onnx to False')
if load_onnx:
self.model.load_onnx('{}/flow.decoder.estimator.fp32.onnx'.format(model_dir))
if load_trt:
self.model.load_trt('{}/flow.decoder.estimator.fp16.Volta.plan'.format(model_dir))
del configs