deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-05 01:49:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

add stream code

Browse Source
This commit is contained in:
lyuxiang.lx
2024-07-23 00:02:30 +08:00
parent 2895d99b9a
commit a13411c561
4 changed files with 123 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
cosyvoice/cli/cosyvoice.py
View File

@@ -12,11 +12,12 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import torch
import time
from hyperpyyaml import load_hyperpyyaml
from modelscope import snapshot_download
from cosyvoice.cli.frontend import CosyVoiceFrontEnd
from cosyvoice.cli.model import CosyVoiceModel
from cosyvoice.utils.file_utils import logging
class CosyVoice:
@@ -44,40 +45,48 @@ class CosyVoice:
spks = list(self.frontend.spk2info.keys())
return spks
def inference_sft(self, tts_text, spk_id):
tts_speeches = []
def inference_sft(self, tts_text, spk_id, stream=False):
start_time = time.time()
for i in self.frontend.text_normalize(tts_text, split=True):
model_input = self.frontend.frontend_sft(i, spk_id)
model_output = self.model.inference(**model_input)
tts_speeches.append(model_output['tts_speech'])
return {'tts_speech': torch.concat(tts_speeches, dim=1)}
for model_output in self.model.inference(**model_input, stream=stream):
speech_len = model_output['tts_speech'].shape[1] / 22050
logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
yield model_output
start_time = time.time()
def inference_zero_shot(self, tts_text, prompt_text, prompt_speech_16k):
def inference_zero_shot(self, tts_text, prompt_text, prompt_speech_16k, stream=False):
start_time = time.time()
prompt_text = self.frontend.text_normalize(prompt_text, split=False)
tts_speeches = []
for i in self.frontend.text_normalize(tts_text, split=True):
model_input = self.frontend.frontend_zero_shot(i, prompt_text, prompt_speech_16k)
model_output = self.model.inference(**model_input)
tts_speeches.append(model_output['tts_speech'])
return {'tts_speech': torch.concat(tts_speeches, dim=1)}
for model_output in self.model.inference(**model_input, stream=stream):
speech_len = model_output['tts_speech'].shape[1] / 22050
logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
yield model_output
start_time = time.time()
def inference_cross_lingual(self, tts_text, prompt_speech_16k):
def inference_cross_lingual(self, tts_text, prompt_speech_16k, stream=False):
if self.frontend.instruct is True:
raise ValueError('{} do not support cross_lingual inference'.format(self.model_dir))
tts_speeches = []
start_time = time.time()
for i in self.frontend.text_normalize(tts_text, split=True):
model_input = self.frontend.frontend_cross_lingual(i, prompt_speech_16k)
model_output = self.model.inference(**model_input)
tts_speeches.append(model_output['tts_speech'])
return {'tts_speech': torch.concat(tts_speeches, dim=1)}
for model_output in self.model.inference(**model_input, stream=stream):
speech_len = model_output['tts_speech'].shape[1] / 22050
logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
yield model_output
start_time = time.time()
def inference_instruct(self, tts_text, spk_id, instruct_text):
def inference_instruct(self, tts_text, spk_id, instruct_text, stream=False):
if self.frontend.instruct is False:
raise ValueError('{} do not support instruct inference'.format(self.model_dir))
start_time = time.time()
instruct_text = self.frontend.text_normalize(instruct_text, split=False)
tts_speeches = []
for i in self.frontend.text_normalize(tts_text, split=True):
model_input = self.frontend.frontend_instruct(i, spk_id, instruct_text)
model_output = self.model.inference(**model_input)
tts_speeches.append(model_output['tts_speech'])
return {'tts_speech': torch.concat(tts_speeches, dim=1)}
for model_output in self.model.inference(**model_input, stream=stream):
speech_len = model_output['tts_speech'].shape[1] / 22050
logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
yield model_output
start_time = time.time()