deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-04 17:39:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

update stream code

Browse Source
This commit is contained in:
lyuxiang.lx
2024-07-30 16:11:28 +08:00
parent 02f941d348
commit f4e70e222c
15 changed files with 182 additions and 109 deletions
Download Patch File Download Diff File Expand all files Collapse all files

133
cosyvoice/cli/model.py
View File

@@ -16,6 +16,8 @@ import numpy as np
import threading
import time
from contextlib import nullcontext
import uuid
from cosyvoice.utils.common import fade_in_out
class CosyVoiceModel:
@@ -28,13 +30,19 @@ class CosyVoiceModel:
self.llm = llm
self.flow = flow
self.hift = hift
self.stream_win_len = 60 * 4
self.stream_hop_len = 50 * 4
self.overlap = 4395 * 4 # 10 token equals 4395 sample point
self.window = np.hamming(2 * self.overlap)
self.token_min_hop_len = 100
self.token_max_hop_len = 400
self.token_overlap_len = 20
self.speech_overlap_len = 34 * 256
self.window = np.hamming(2 * self.speech_overlap_len)
self.stream_scale_factor = 1
assert self.stream_scale_factor >= 1, 'stream_scale_factor should be greater than 1, change it according to your actual rtf'
self.llm_context = torch.cuda.stream(torch.cuda.Stream(self.device)) if torch.cuda.is_available() else nullcontext()
self.flow_hift_context = torch.cuda.stream(torch.cuda.Stream(self.device)) if torch.cuda.is_available() else nullcontext()
self.lock = threading.Lock()
# dict used to store session related variable
self.tts_speech_token = {}
self.llm_end = {}
def load(self, llm_model, flow_model, hift_model):
self.llm.load_state_dict(torch.load(llm_model, map_location=self.device))
@@ -44,7 +52,7 @@ class CosyVoiceModel:
self.hift.load_state_dict(torch.load(hift_model, map_location=self.device))
self.hift.to(self.device).eval()
def llm_job(self, text, text_len, prompt_text, prompt_text_len, llm_prompt_speech_token, llm_prompt_speech_token_len, llm_embedding):
def llm_job(self, text, text_len, prompt_text, prompt_text_len, llm_prompt_speech_token, llm_prompt_speech_token_len, llm_embedding, this_uuid):
with self.llm_context:
for i in self.llm.inference(text=text.to(self.device),
text_len=text_len.to(self.device),
@@ -53,13 +61,11 @@ class CosyVoiceModel:
prompt_speech_token=llm_prompt_speech_token.to(self.device),
prompt_speech_token_len=llm_prompt_speech_token_len.to(self.device),
embedding=llm_embedding.to(self.device),
beam_size=1,
sampling=25,
max_token_text_ratio=30,
min_token_text_ratio=3,
stream=True):
self.tts_speech_token.append(i)
self.llm_end = True
min_token_text_ratio=3):
self.tts_speech_token[this_uuid].append(i)
self.llm_end[this_uuid] = True
def token2wav(self, token, prompt_token, prompt_token_len, prompt_feat, prompt_feat_len, embedding):
with self.flow_hift_context:
@@ -78,15 +84,19 @@ class CosyVoiceModel:
llm_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32), llm_prompt_speech_token_len=torch.zeros(1, dtype=torch.int32),
flow_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32), flow_prompt_speech_token_len=torch.zeros(1, dtype=torch.int32),
prompt_speech_feat=torch.zeros(1, 0, 80), prompt_speech_feat_len=torch.zeros(1, dtype=torch.int32), stream=False):
# this_uuid is used to track variables related to this inference thread
this_uuid = str(uuid.uuid1())
with self.lock:
self.tts_speech_token[this_uuid], self.llm_end[this_uuid] = [], False
p = threading.Thread(target=self.llm_job, args=(text.to(self.device), text_len.to(self.device), prompt_text.to(self.device), prompt_text_len.to(self.device),
llm_prompt_speech_token.to(self.device), llm_prompt_speech_token_len.to(self.device), llm_embedding.to(self.device), this_uuid))
p.start()
if stream is True:
self.tts_speech_token, self.llm_end, cache_speech = [], False, None
p = threading.Thread(target=self.llm_job, args=(text.to(self.device), text_len.to(self.device), prompt_text.to(self.device), prompt_text_len.to(self.device),
llm_prompt_speech_token.to(self.device), llm_prompt_speech_token_len.to(self.device), llm_embedding.to(self.device)))
p.start()
cache_speech, cache_token, token_hop_len = None, None, self.token_min_hop_len
while True:
time.sleep(0.1)
if len(self.tts_speech_token) >= self.stream_win_len:
this_tts_speech_token = torch.concat(self.tts_speech_token[:self.stream_win_len], dim=1)
if len(self.tts_speech_token[this_uuid]) >= token_hop_len + self.token_overlap_len:
this_tts_speech_token = torch.concat(self.tts_speech_token[this_uuid][:token_hop_len + self.token_overlap_len], dim=1)
with self.flow_hift_context:
this_tts_speech = self.token2wav(token=this_tts_speech_token,
prompt_token=flow_prompt_speech_token.to(self.device),
@@ -96,57 +106,48 @@ class CosyVoiceModel:
embedding=flow_embedding.to(self.device))
# fade in/out if necessary
if cache_speech is not None:
this_tts_speech[:, :self.overlap] = this_tts_speech[:, :self.overlap] * self.window[:self.overlap] + cache_speech * self.window[-self.overlap:]
yield {'tts_speech': this_tts_speech[:, :-self.overlap]}
cache_speech = this_tts_speech[:, -self.overlap:]
this_tts_speech = fade_in_out(this_tts_speech, cache_speech, self.window)
yield {'tts_speech': this_tts_speech[:, :-self.speech_overlap_len]}
cache_speech = this_tts_speech[:, -self.speech_overlap_len:]
cache_token = self.tts_speech_token[this_uuid][:token_hop_len]
with self.lock:
self.tts_speech_token = self.tts_speech_token[self.stream_hop_len:]
if self.llm_end is True:
self.tts_speech_token[this_uuid] = self.tts_speech_token[this_uuid][token_hop_len:]
# increase token_hop_len for better speech quality
token_hop_len = min(self.token_max_hop_len, int(token_hop_len * self.stream_scale_factor))
if self.llm_end[this_uuid] is True and len(self.tts_speech_token[this_uuid]) < token_hop_len + self.token_overlap_len:
break
# deal with remain tokens
if cache_speech is None or len(self.tts_speech_token) > self.stream_win_len - self.stream_hop_len:
this_tts_speech_token = torch.concat(self.tts_speech_token, dim=1)
with self.flow_hift_context:
this_tts_mel = self.flow.inference(token=this_tts_speech_token,
token_len=torch.tensor([this_tts_speech_token.size(1)], dtype=torch.int32).to(self.device),
prompt_token=flow_prompt_speech_token.to(self.device),
prompt_token_len=flow_prompt_speech_token_len.to(self.device),
prompt_feat=prompt_speech_feat.to(self.device),
prompt_feat_len=prompt_speech_feat_len.to(self.device),
embedding=flow_embedding.to(self.device))
this_tts_speech = self.hift.inference(mel=this_tts_mel).cpu()
if cache_speech is not None:
this_tts_speech[:, :self.overlap] = this_tts_speech[:, :self.overlap] * self.window[:self.overlap] + cache_speech * self.window[-self.overlap:]
yield {'tts_speech': this_tts_speech}
else:
assert len(self.tts_speech_token) == self.stream_win_len - self.stream_hop_len, 'tts_speech_token not equal to {}'.format(self.stream_win_len - self.stream_hop_len)
yield {'tts_speech': cache_speech}
p.join()
torch.cuda.synchronize()
# deal with remain tokens, make sure inference remain token len equals token_hop_len when cache_speech is not None
this_tts_speech_token = torch.concat(self.tts_speech_token[this_uuid], dim=1)
if this_tts_speech_token.shape[1] < self.token_min_hop_len + self.token_overlap_len and cache_token is not None:
cache_token_len = self.token_min_hop_len + self.token_overlap_len - this_tts_speech_token.shape[1]
this_tts_speech_token = torch.concat([torch.concat(cache_token[-cache_token_len:], dim=1), this_tts_speech_token], dim=1)
else:
cache_token_len = 0
with self.flow_hift_context:
this_tts_speech = self.token2wav(token=this_tts_speech_token,
prompt_token=flow_prompt_speech_token.to(self.device),
prompt_token_len=flow_prompt_speech_token_len.to(self.device),
prompt_feat=prompt_speech_feat.to(self.device),
prompt_feat_len=prompt_speech_feat_len.to(self.device),
embedding=flow_embedding.to(self.device))
this_tts_speech = this_tts_speech[:, int(cache_token_len / this_tts_speech_token.shape[1] * this_tts_speech.shape[1]):]
if cache_speech is not None:
this_tts_speech = fade_in_out(this_tts_speech, cache_speech, self.window)
yield {'tts_speech': this_tts_speech}
else:
tts_speech_token = []
for i in self.llm.inference(text=text.to(self.device),
text_len=text_len.to(self.device),
prompt_text=prompt_text.to(self.device),
prompt_text_len=prompt_text_len.to(self.device),
prompt_speech_token=llm_prompt_speech_token.to(self.device),
prompt_speech_token_len=llm_prompt_speech_token_len.to(self.device),
embedding=llm_embedding.to(self.device),
beam_size=1,
sampling=25,
max_token_text_ratio=30,
min_token_text_ratio=3,
stream=stream):
tts_speech_token.append(i)
assert len(tts_speech_token) == 1, 'tts_speech_token len should be 1 when stream is {}'.format(stream)
tts_speech_token = torch.concat(tts_speech_token, dim=1)
tts_mel = self.flow.inference(token=tts_speech_token,
token_len=torch.tensor([tts_speech_token.size(1)], dtype=torch.int32).to(self.device),
prompt_token=flow_prompt_speech_token.to(self.device),
prompt_token_len=flow_prompt_speech_token_len.to(self.device),
prompt_feat=prompt_speech_feat.to(self.device),
prompt_feat_len=prompt_speech_feat_len.to(self.device),
embedding=flow_embedding.to(self.device))
tts_speech = self.hift.inference(mel=tts_mel).cpu()
torch.cuda.empty_cache()
yield {'tts_speech': tts_speech}
# deal with all tokens
p.join()
this_tts_speech_token = torch.concat(self.tts_speech_token[this_uuid], dim=1)
with self.flow_hift_context:
this_tts_speech = self.token2wav(token=this_tts_speech_token,
prompt_token=flow_prompt_speech_token.to(self.device),
prompt_token_len=flow_prompt_speech_token_len.to(self.device),
prompt_feat=prompt_speech_feat.to(self.device),
prompt_feat_len=prompt_speech_feat_len.to(self.device),
embedding=flow_embedding.to(self.device))
yield {'tts_speech': this_tts_speech}
with self.lock:
self.tts_speech_token.pop(this_uuid)
self.llm_end.pop(this_uuid)
torch.cuda.synchronize()