liuhua
liuhua
Kevin Hu
commited on
Commit
·
32ef5e5
1
Parent(s):
45b5929
OpenAITTS (#2493)
Browse files### What problem does this PR solve?
OpenAITTS
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
---------
Co-authored-by: liuhua <[email protected]>
Co-authored-by: Kevin Hu <[email protected]>
- conf/llm_factories.json +6 -0
- rag/llm/__init__.py +2 -1
- rag/llm/tts_model.py +46 -16
conf/llm_factories.json
CHANGED
@@ -77,6 +77,12 @@
|
|
77 |
"tags": "LLM,CHAT,IMAGE2TEXT",
|
78 |
"max_tokens": 765,
|
79 |
"model_type": "image2text"
|
|
|
|
|
|
|
|
|
|
|
|
|
80 |
}
|
81 |
]
|
82 |
},
|
|
|
77 |
"tags": "LLM,CHAT,IMAGE2TEXT",
|
78 |
"max_tokens": 765,
|
79 |
"model_type": "image2text"
|
80 |
+
},
|
81 |
+
{
|
82 |
+
"llm_name": "tts-1",
|
83 |
+
"tags": "TTS",
|
84 |
+
"max_tokens": 2048,
|
85 |
+
"model_type": "tts"
|
86 |
}
|
87 |
]
|
88 |
},
|
rag/llm/__init__.py
CHANGED
@@ -138,5 +138,6 @@ Seq2txtModel = {
|
|
138 |
|
139 |
TTSModel = {
|
140 |
"Fish Audio": FishAudioTTS,
|
141 |
-
"Tongyi-Qianwen": QwenTTS
|
|
|
142 |
}
|
|
|
138 |
|
139 |
TTSModel = {
|
140 |
"Fish Audio": FishAudioTTS,
|
141 |
+
"Tongyi-Qianwen": QwenTTS,
|
142 |
+
"OpenAI":OpenAITTS
|
143 |
}
|
rag/llm/tts_model.py
CHANGED
@@ -14,6 +14,7 @@
|
|
14 |
# limitations under the License.
|
15 |
#
|
16 |
|
|
|
17 |
from typing import Annotated, Literal
|
18 |
from abc import ABC
|
19 |
import httpx
|
@@ -23,6 +24,8 @@ from rag.utils import num_tokens_from_string
|
|
23 |
import json
|
24 |
import re
|
25 |
import time
|
|
|
|
|
26 |
class ServeReferenceAudio(BaseModel):
|
27 |
audio: bytes
|
28 |
text: str
|
@@ -52,7 +55,7 @@ class Base(ABC):
|
|
52 |
|
53 |
def tts(self, audio):
|
54 |
pass
|
55 |
-
|
56 |
def normalize_text(self, text):
|
57 |
return re.sub(r'(\*\*|##\d+\$\$|#)', '', text)
|
58 |
|
@@ -78,13 +81,13 @@ class FishAudioTTS(Base):
|
|
78 |
with httpx.Client() as client:
|
79 |
try:
|
80 |
with client.stream(
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
) as response:
|
89 |
if response.status_code == HTTPStatus.OK:
|
90 |
for chunk in response.iter_bytes():
|
@@ -101,7 +104,7 @@ class FishAudioTTS(Base):
|
|
101 |
class QwenTTS(Base):
|
102 |
def __init__(self, key, model_name, base_url=""):
|
103 |
import dashscope
|
104 |
-
|
105 |
self.model_name = model_name
|
106 |
dashscope.api_key = key
|
107 |
|
@@ -109,11 +112,11 @@ class QwenTTS(Base):
|
|
109 |
from dashscope.api_entities.dashscope_response import SpeechSynthesisResponse
|
110 |
from dashscope.audio.tts import ResultCallback, SpeechSynthesizer, SpeechSynthesisResult
|
111 |
from collections import deque
|
112 |
-
|
113 |
class Callback(ResultCallback):
|
114 |
def __init__(self) -> None:
|
115 |
self.dque = deque()
|
116 |
-
|
117 |
def _run(self):
|
118 |
while True:
|
119 |
if not self.dque:
|
@@ -144,13 +147,40 @@ class QwenTTS(Base):
|
|
144 |
text = self.normalize_text(text)
|
145 |
callback = Callback()
|
146 |
SpeechSynthesizer.call(model=self.model_name,
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
try:
|
151 |
for data in callback._run():
|
152 |
yield data
|
153 |
yield num_tokens_from_string(text)
|
154 |
-
|
155 |
except Exception as e:
|
156 |
-
raise RuntimeError(f"**ERROR**: {e}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
# limitations under the License.
|
15 |
#
|
16 |
|
17 |
+
import requests
|
18 |
from typing import Annotated, Literal
|
19 |
from abc import ABC
|
20 |
import httpx
|
|
|
24 |
import json
|
25 |
import re
|
26 |
import time
|
27 |
+
|
28 |
+
|
29 |
class ServeReferenceAudio(BaseModel):
|
30 |
audio: bytes
|
31 |
text: str
|
|
|
55 |
|
56 |
def tts(self, audio):
|
57 |
pass
|
58 |
+
|
59 |
def normalize_text(self, text):
|
60 |
return re.sub(r'(\*\*|##\d+\$\$|#)', '', text)
|
61 |
|
|
|
81 |
with httpx.Client() as client:
|
82 |
try:
|
83 |
with client.stream(
|
84 |
+
method="POST",
|
85 |
+
url=self.base_url,
|
86 |
+
content=ormsgpack.packb(
|
87 |
+
request, option=ormsgpack.OPT_SERIALIZE_PYDANTIC
|
88 |
+
),
|
89 |
+
headers=self.headers,
|
90 |
+
timeout=None,
|
91 |
) as response:
|
92 |
if response.status_code == HTTPStatus.OK:
|
93 |
for chunk in response.iter_bytes():
|
|
|
104 |
class QwenTTS(Base):
|
105 |
def __init__(self, key, model_name, base_url=""):
|
106 |
import dashscope
|
107 |
+
|
108 |
self.model_name = model_name
|
109 |
dashscope.api_key = key
|
110 |
|
|
|
112 |
from dashscope.api_entities.dashscope_response import SpeechSynthesisResponse
|
113 |
from dashscope.audio.tts import ResultCallback, SpeechSynthesizer, SpeechSynthesisResult
|
114 |
from collections import deque
|
115 |
+
|
116 |
class Callback(ResultCallback):
|
117 |
def __init__(self) -> None:
|
118 |
self.dque = deque()
|
119 |
+
|
120 |
def _run(self):
|
121 |
while True:
|
122 |
if not self.dque:
|
|
|
147 |
text = self.normalize_text(text)
|
148 |
callback = Callback()
|
149 |
SpeechSynthesizer.call(model=self.model_name,
|
150 |
+
text=text,
|
151 |
+
callback=callback,
|
152 |
+
format="mp3")
|
153 |
try:
|
154 |
for data in callback._run():
|
155 |
yield data
|
156 |
yield num_tokens_from_string(text)
|
157 |
+
|
158 |
except Exception as e:
|
159 |
+
raise RuntimeError(f"**ERROR**: {e}")
|
160 |
+
|
161 |
+
|
162 |
+
class OpenAITTS(Base):
|
163 |
+
def __init__(self, key, model_name="tts-1", base_url="https://api.openai.com/v1"):
|
164 |
+
self.api_key = key
|
165 |
+
self.model_name = model_name
|
166 |
+
self.base_url = base_url
|
167 |
+
self.headers = {
|
168 |
+
"Authorization": f"Bearer {self.api_key}",
|
169 |
+
"Content-Type": "application/json"
|
170 |
+
}
|
171 |
+
|
172 |
+
def tts(self, text, voice="alloy"):
|
173 |
+
text = self.normalize_text(text)
|
174 |
+
payload = {
|
175 |
+
"model": self.model_name,
|
176 |
+
"voice": voice,
|
177 |
+
"input": text
|
178 |
+
}
|
179 |
+
|
180 |
+
response = requests.post(f"{self.base_url}/audio/speech", headers=self.headers, json=payload, stream=True)
|
181 |
+
|
182 |
+
if response.status_code != 200:
|
183 |
+
raise Exception(f"**Error**: {response.status_code}, {response.text}")
|
184 |
+
for chunk in response.iter_content(chunk_size=1024):
|
185 |
+
if chunk:
|
186 |
+
yield chunk
|