Create orpheus_engine.py
Browse files- orpheus_engine.py +373 -0
orpheus_engine.py
ADDED
@@ -0,0 +1,373 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import json
|
2 |
+
import time
|
3 |
+
import logging
|
4 |
+
import pyaudio
|
5 |
+
import requests
|
6 |
+
import traceback
|
7 |
+
import numpy as np
|
8 |
+
from queue import Queue
|
9 |
+
from typing import Optional, Union
|
10 |
+
from .base_engine import BaseEngine
|
11 |
+
|
12 |
+
# Default configuration values
|
13 |
+
DEFAULT_API_URL = "http://127.0.0.1:1234/v1/completions"
|
14 |
+
DEFAULT_HEADERS = {"Content-Type": "application/json"}
|
15 |
+
DEFAULT_MODEL = "orpheus-3b-0.1-ft"
|
16 |
+
DEFAULT_VOICE = "tara"
|
17 |
+
AVAILABLE_VOICES = ["tara", "leah", "jess", "leo", "dan", "mia", "zac", "zoe"]
|
18 |
+
SAMPLE_RATE = 24000 # Specific sample rate for Orpheus
|
19 |
+
|
20 |
+
# Special token definitions for prompt formatting and token decoding
|
21 |
+
START_TOKEN_ID = 128259
|
22 |
+
END_TOKEN_IDS = [128009, 128260, 128261, 128257]
|
23 |
+
CUSTOM_TOKEN_PREFIX = "<custom_token_"
|
24 |
+
|
25 |
+
|
26 |
+
class OrpheusVoice:
|
27 |
+
"""
|
28 |
+
Represents the configuration for an Orpheus voice.
|
29 |
+
|
30 |
+
Attributes:
|
31 |
+
name (str): The name of the voice. Must be one of the AVAILABLE_VOICES.
|
32 |
+
|
33 |
+
Raises:
|
34 |
+
ValueError: If the voice name provided is not in AVAILABLE_VOICES.
|
35 |
+
"""
|
36 |
+
def __init__(self, name: str):
|
37 |
+
# if name not in AVAILABLE_VOICES:
|
38 |
+
# raise ValueError(f"Invalid voice '{name}'. Available voices: {AVAILABLE_VOICES}")
|
39 |
+
self.name = name
|
40 |
+
|
41 |
+
def __repr__(self):
|
42 |
+
return f"OrpheusVoice(name='{self.name}')"
|
43 |
+
|
44 |
+
|
45 |
+
class OrpheusEngine(BaseEngine):
|
46 |
+
"""
|
47 |
+
Real-time Text-to-Speech (TTS) engine for the Orpheus model via LM Studio API.
|
48 |
+
|
49 |
+
This engine supports real-time token generation, audio synthesis, and voice configuration.
|
50 |
+
"""
|
51 |
+
|
52 |
+
def __init__(
|
53 |
+
self,
|
54 |
+
api_url: str = DEFAULT_API_URL,
|
55 |
+
model: str = DEFAULT_MODEL,
|
56 |
+
headers: dict = DEFAULT_HEADERS,
|
57 |
+
voice: Optional[OrpheusVoice] = None,
|
58 |
+
temperature: float = 0.6,
|
59 |
+
top_p: float = 0.9,
|
60 |
+
max_tokens: int = 1200,
|
61 |
+
repetition_penalty: float = 1.1,
|
62 |
+
debug: bool = False
|
63 |
+
):
|
64 |
+
"""
|
65 |
+
Initialize the Orpheus TTS engine with the given parameters.
|
66 |
+
|
67 |
+
Args:
|
68 |
+
api_url (str): Endpoint URL for the LM Studio API.
|
69 |
+
model (str): Model name to use for synthesis.
|
70 |
+
headers (dict): HTTP headers for API requests.
|
71 |
+
voice (Optional[OrpheusVoice]): OrpheusVoice configuration. Defaults to DEFAULT_VOICE.
|
72 |
+
temperature (float): Sampling temperature (0-1) for text generation.
|
73 |
+
top_p (float): Top-p sampling parameter for controlling diversity.
|
74 |
+
max_tokens (int): Maximum tokens to generate per API request.
|
75 |
+
repetition_penalty (float): Penalty factor for repeated phrases.
|
76 |
+
debug (bool): Flag to enable debug output.
|
77 |
+
"""
|
78 |
+
super().__init__()
|
79 |
+
self.api_url = api_url
|
80 |
+
self.model = model
|
81 |
+
self.headers = headers
|
82 |
+
self.voice = voice or OrpheusVoice(DEFAULT_VOICE)
|
83 |
+
self.temperature = temperature
|
84 |
+
self.top_p = top_p
|
85 |
+
self.max_tokens = max_tokens
|
86 |
+
self.repetition_penalty = repetition_penalty
|
87 |
+
self.debug = debug
|
88 |
+
self.queue = Queue()
|
89 |
+
self.post_init()
|
90 |
+
|
91 |
+
def post_init(self):
|
92 |
+
"""Set up additional engine attributes."""
|
93 |
+
self.engine_name = "orpheus"
|
94 |
+
|
95 |
+
def get_stream_info(self):
|
96 |
+
"""
|
97 |
+
Retrieve PyAudio stream configuration.
|
98 |
+
|
99 |
+
Returns:
|
100 |
+
tuple: Format, channel count, and sample rate for PyAudio.
|
101 |
+
"""
|
102 |
+
return pyaudio.paInt16, 1, SAMPLE_RATE
|
103 |
+
|
104 |
+
def synthesize(self, text: str) -> bool:
|
105 |
+
"""
|
106 |
+
Convert text to speech and stream audio data.
|
107 |
+
|
108 |
+
Args:
|
109 |
+
text (str): The input text to be synthesized.
|
110 |
+
|
111 |
+
Returns:
|
112 |
+
bool: True if synthesis was successful, False otherwise.
|
113 |
+
"""
|
114 |
+
super().synthesize(text)
|
115 |
+
|
116 |
+
try:
|
117 |
+
# Process tokens and put generated audio chunks into the queue
|
118 |
+
for audio_chunk in self._token_decoder(self._generate_tokens(text)):
|
119 |
+
# bail out immediately if someone called .stop()
|
120 |
+
if self.stop_synthesis_event.is_set():
|
121 |
+
logging.info("OrpheusEngine: synthesis stopped by user")
|
122 |
+
return False
|
123 |
+
print(f"Audio chunk size: {len(audio_chunk)}")
|
124 |
+
self.queue.put(audio_chunk)
|
125 |
+
return True
|
126 |
+
except Exception as e:
|
127 |
+
traceback.print_exc()
|
128 |
+
logging.error(f"Synthesis error: {e}")
|
129 |
+
return False
|
130 |
+
|
131 |
+
def synthesize(self, text: str) -> bool:
|
132 |
+
"""
|
133 |
+
Convert text to speech and stream audio data via Orpheus.
|
134 |
+
Drops initial and trailing near-silent chunks.
|
135 |
+
"""
|
136 |
+
super().synthesize(text)
|
137 |
+
|
138 |
+
try:
|
139 |
+
for audio_chunk in self._token_decoder(self._generate_tokens(text)):
|
140 |
+
# bail out if user called .stop()
|
141 |
+
if self.stop_synthesis_event.is_set():
|
142 |
+
logging.info("OrpheusEngine: synthesis stopped by user")
|
143 |
+
return False
|
144 |
+
|
145 |
+
# forward this chunk
|
146 |
+
self.queue.put(audio_chunk)
|
147 |
+
|
148 |
+
return True
|
149 |
+
|
150 |
+
except Exception as e:
|
151 |
+
traceback.print_exc()
|
152 |
+
logging.error(f"Synthesis error: {e}")
|
153 |
+
return False
|
154 |
+
|
155 |
+
|
156 |
+
def _generate_tokens(self, prompt: str):
|
157 |
+
"""
|
158 |
+
Generate a token stream using the LM Studio API.
|
159 |
+
|
160 |
+
Args:
|
161 |
+
prompt (str): The input text prompt.
|
162 |
+
|
163 |
+
Yields:
|
164 |
+
str: Each token's text as it is received from the API.
|
165 |
+
"""
|
166 |
+
logging.debug(f"Generating tokens for prompt: {prompt}")
|
167 |
+
formatted_prompt = self._format_prompt(prompt)
|
168 |
+
|
169 |
+
payload = {
|
170 |
+
"model": self.model,
|
171 |
+
"prompt": formatted_prompt,
|
172 |
+
"max_tokens": self.max_tokens,
|
173 |
+
"temperature": self.temperature,
|
174 |
+
"top_p": self.top_p,
|
175 |
+
"repeat_penalty": self.repetition_penalty,
|
176 |
+
"stream": True
|
177 |
+
}
|
178 |
+
|
179 |
+
try:
|
180 |
+
logging.debug(f"Requesting API URL: {self.api_url} with payload: {payload} and headers: {self.headers}")
|
181 |
+
response = requests.post(
|
182 |
+
self.api_url,
|
183 |
+
headers=self.headers,
|
184 |
+
json=payload,
|
185 |
+
stream=True
|
186 |
+
)
|
187 |
+
response.raise_for_status()
|
188 |
+
|
189 |
+
token_counter = 0
|
190 |
+
start_time = time.time() # Start timing token generation
|
191 |
+
for line in response.iter_lines():
|
192 |
+
# stop on demand
|
193 |
+
if self.stop_synthesis_event.is_set():
|
194 |
+
logging.debug("OrpheusEngine: token generation aborted")
|
195 |
+
break
|
196 |
+
if line:
|
197 |
+
line = line.decode('utf-8')
|
198 |
+
if line.startswith('data: '):
|
199 |
+
data_str = line[6:]
|
200 |
+
if data_str.strip() == '[DONE]':
|
201 |
+
break
|
202 |
+
|
203 |
+
try:
|
204 |
+
data = json.loads(data_str)
|
205 |
+
if 'choices' in data and data['choices']:
|
206 |
+
token_text = data['choices'][0].get('text', '')
|
207 |
+
if token_text:
|
208 |
+
token_counter += 1
|
209 |
+
# Print the time it took to get the first token
|
210 |
+
if token_counter == 1:
|
211 |
+
elapsed = time.time() - start_time
|
212 |
+
logging.info(f"Time to first token: {elapsed:.2f} seconds")
|
213 |
+
yield token_text
|
214 |
+
except json.JSONDecodeError as e:
|
215 |
+
logging.error(f"Error decoding JSON: {e}")
|
216 |
+
continue
|
217 |
+
|
218 |
+
except requests.RequestException as e:
|
219 |
+
logging.error(f"API request failed: {e}")
|
220 |
+
|
221 |
+
def _format_prompt(self, prompt: str) -> str:
|
222 |
+
"""
|
223 |
+
Format the text prompt with special tokens required by Orpheus.
|
224 |
+
|
225 |
+
Args:
|
226 |
+
prompt (str): The raw text prompt.
|
227 |
+
|
228 |
+
Returns:
|
229 |
+
str: The formatted prompt including voice and termination token.
|
230 |
+
"""
|
231 |
+
return f"<|audio|>{self.voice.name}: {prompt}<|eot_id|>"
|
232 |
+
|
233 |
+
def _token_decoder(self, token_gen):
|
234 |
+
"""
|
235 |
+
Decode tokens from the generator and convert them into audio samples.
|
236 |
+
|
237 |
+
This method aggregates tokens in a buffer and converts them into audio chunks
|
238 |
+
once enough tokens have been collected.
|
239 |
+
|
240 |
+
Args:
|
241 |
+
token_gen: Generator yielding token strings.
|
242 |
+
|
243 |
+
Yields:
|
244 |
+
Audio samples ready to be streamed.
|
245 |
+
"""
|
246 |
+
buffer = []
|
247 |
+
count = 0
|
248 |
+
|
249 |
+
logging.debug("Starting token decoding from token generator.")
|
250 |
+
for token_text in token_gen:
|
251 |
+
# bail out if stop was requested
|
252 |
+
if self.stop_synthesis_event.is_set():
|
253 |
+
logging.debug("OrpheusEngine: token decoding aborted")
|
254 |
+
break
|
255 |
+
token = self.turn_token_into_id(token_text, count)
|
256 |
+
if token is not None and token > 0:
|
257 |
+
buffer.append(token)
|
258 |
+
count += 1
|
259 |
+
|
260 |
+
# Process every 7 tokens after an initial threshold
|
261 |
+
if count % 7 == 0 and count > 27:
|
262 |
+
buffer_to_proc = buffer[-28:]
|
263 |
+
audio_samples = self._convert_buffer(buffer_to_proc, count)
|
264 |
+
if audio_samples is not None:
|
265 |
+
yield audio_samples
|
266 |
+
|
267 |
+
def turn_token_into_id(self, token_string: str, index: int) -> Optional[int]:
|
268 |
+
"""
|
269 |
+
Convert a token string to a numeric ID for audio processing.
|
270 |
+
|
271 |
+
The conversion takes into account the custom token prefix and an index-based offset.
|
272 |
+
|
273 |
+
Args:
|
274 |
+
token_string (str): The token text.
|
275 |
+
index (int): The current token index.
|
276 |
+
|
277 |
+
Returns:
|
278 |
+
Optional[int]: The numeric token ID or None if conversion fails.
|
279 |
+
"""
|
280 |
+
token_string = token_string.strip()
|
281 |
+
last_token_start = token_string.rfind(CUSTOM_TOKEN_PREFIX)
|
282 |
+
|
283 |
+
if last_token_start == -1:
|
284 |
+
return None
|
285 |
+
|
286 |
+
last_token = token_string[last_token_start:]
|
287 |
+
|
288 |
+
if last_token.startswith(CUSTOM_TOKEN_PREFIX) and last_token.endswith(">"):
|
289 |
+
try:
|
290 |
+
number_str = last_token[14:-1]
|
291 |
+
token_id = int(number_str) - 10 - ((index % 7) * 4096)
|
292 |
+
return token_id
|
293 |
+
except ValueError:
|
294 |
+
return None
|
295 |
+
else:
|
296 |
+
return None
|
297 |
+
|
298 |
+
def _convert_buffer(self, multiframe, count: int):
|
299 |
+
"""
|
300 |
+
Convert a buffer of token frames into audio samples.
|
301 |
+
|
302 |
+
This method uses an external decoder to convert the collected token frames.
|
303 |
+
|
304 |
+
Args:
|
305 |
+
multiframe: List of token IDs to be converted.
|
306 |
+
count (int): The current token count (used for conversion logic).
|
307 |
+
|
308 |
+
Returns:
|
309 |
+
Converted audio samples if successful; otherwise, None.
|
310 |
+
"""
|
311 |
+
try:
|
312 |
+
from .orpheus_decoder import convert_to_audio as orpheus_convert_to_audio
|
313 |
+
converted = orpheus_convert_to_audio(multiframe, count)
|
314 |
+
if converted is None:
|
315 |
+
logging.warning("Conversion returned None.")
|
316 |
+
return converted
|
317 |
+
except Exception as e:
|
318 |
+
logging.error(f"Failed to convert buffer to audio: {e}")
|
319 |
+
logging.info("Returning None after failed conversion.")
|
320 |
+
return None
|
321 |
+
|
322 |
+
def get_voices(self):
|
323 |
+
"""
|
324 |
+
Retrieve the list of available voices.
|
325 |
+
|
326 |
+
Returns:
|
327 |
+
list: A list of OrpheusVoice instances for each available voice.
|
328 |
+
"""
|
329 |
+
return [OrpheusVoice(name) for name in AVAILABLE_VOICES]
|
330 |
+
|
331 |
+
def set_voice(self, voice: Union[str, OrpheusVoice]):
|
332 |
+
"""
|
333 |
+
Set the current voice for synthesis.
|
334 |
+
|
335 |
+
Args:
|
336 |
+
voice (Union[str, OrpheusVoice]): The voice name or an OrpheusVoice instance.
|
337 |
+
|
338 |
+
Raises:
|
339 |
+
ValueError: If the provided voice name is invalid.
|
340 |
+
TypeError: If the voice argument is neither a string nor an OrpheusVoice instance.
|
341 |
+
"""
|
342 |
+
if isinstance(voice, str):
|
343 |
+
# if voice not in AVAILABLE_VOICES:
|
344 |
+
# raise ValueError(f"Invalid voice '{voice}'")
|
345 |
+
self.voice = OrpheusVoice(voice)
|
346 |
+
elif isinstance(voice, OrpheusVoice):
|
347 |
+
self.voice = voice
|
348 |
+
else:
|
349 |
+
raise TypeError("Voice must be a string or an OrpheusVoice instance.")
|
350 |
+
|
351 |
+
def set_voice_parameters(self, **kwargs):
|
352 |
+
"""
|
353 |
+
Update voice generation parameters.
|
354 |
+
|
355 |
+
Valid parameters include 'temperature', 'top_p', 'max_tokens', and 'repetition_penalty'.
|
356 |
+
|
357 |
+
Args:
|
358 |
+
**kwargs: Arbitrary keyword arguments for valid voice parameters.
|
359 |
+
"""
|
360 |
+
valid_params = ['temperature', 'top_p', 'max_tokens', 'repetition_penalty']
|
361 |
+
for param, value in kwargs.items():
|
362 |
+
if param in valid_params:
|
363 |
+
setattr(self, param, value)
|
364 |
+
elif self.debug:
|
365 |
+
logging.warning(f"Ignoring invalid parameter: {param}")
|
366 |
+
|
367 |
+
def __del__(self):
|
368 |
+
"""
|
369 |
+
Destructor to clean up resources.
|
370 |
+
|
371 |
+
Puts a None into the queue to signal termination of audio processing.
|
372 |
+
"""
|
373 |
+
self.queue.put(None)
|