Spaces:
Sleeping
Sleeping
File size: 56,120 Bytes
e85cd49 eafd55a e85cd49 eafd55a e85cd49 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 |
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Unified TTS Processor - A Comprehensive Text-to-Speech System
OVERVIEW:
This module provides a unified interface for advanced text-to-speech synthesis using
the BATONTTS model architecture. It supports two distinct inference modes that offer
different levels of control over speech generation.
ARCHITECTURE:
- Main Model: BATONTTS-1.7B (based on Qwen3-1.7B-instruct)
- Audio Generator: CosyVoice2-0.5B for speech token to waveform conversion
- Inference Engine: vLLM for efficient large language model inference
- Token Processing: Custom tokenization and offset handling for speech tokens
TWO INFERENCE MODES:
Mode 1 - Text-to-Speech (Automatic Feature Generation):
Purpose: Generate speech from text input only
Input: Text string
Process: Text → Model generates word_features + speech_tokens → Audio
Use Case: Simple TTS when you don't need fine-grained prosodic control
Implementation: Matches mode1.py exactly for consistency
Mode 2 - Text+Features-to-Speech (Manual Feature Control):
Purpose: Generate speech with explicit prosodic control
Input: Text string + word_features (JSON)
Process: Text + Features → Model generates speech_tokens → Audio
Use Case: Advanced TTS with precise control over pitch, duration, etc.
Implementation: Matches mode2.py exactly for consistency
KEY FEATURES:
- Lazy model loading for memory efficiency
- Voice cloning support via prompt audio
- Configurable speech speed control
- Comprehensive error handling and logging
- GPU memory optimization
- Tensor parallelism support for multi-GPU setups
USAGE PATTERNS:
# Basic text-to-speech
tts = UnifiedTTS()
tts.text_to_speech("Hello world", "output.wav")
# Advanced prosodic control
features = '[{"word": "Hello", "pitch_mean": 300, "duration": 0.5}]'
tts.text_features_to_speech("Hello", features, "output.wav")
# Cleanup when done
tts.cleanup()
"""
import os
import sys
import json
import torch
import torchaudio
import numpy as np
import argparse
import uuid
import logging
from typing import List, Optional, Tuple, Dict, Any
from transformers import AutoTokenizer
from vllm import LLM, SamplingParams
from modelscope import snapshot_download
cosyvoice_cache_dir = snapshot_download('iic/CosyVoice2-0.5B')
# Add CosyVoice paths
sys.path.append('third-party/CosyVoice')
sys.path.append('third-party/Matcha-TTS')
from cosyvoice.cli.cosyvoice import CosyVoice2
from cosyvoice.utils.file_utils import load_wav
# Configure logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)
class UnifiedTTS:
"""
Unified Text-to-Speech Processor
This class provides a unified interface for two TTS modes:
1. Mode 1 (text_to_speech): Input text only, generates word_features and speech_token internally
2. Mode 2 (text_features_to_speech): Input text and word_features, generates speech_token only
Architecture:
- Uses vLLM for efficient model inference
- Integrates CosyVoice2 for speech token to audio conversion
- Supports both emotion-controlled and standard TTS generation
- Provides easy-to-use API for external script integration
Usage Example:
tts = UnifiedTTS()
# Mode 1: Text only
success = tts.text_to_speech("Hello world", "output1.wav")
# Mode 2: Text + features
features = '[{"word": "Hello", "pitch_mean": 300}]'
success = tts.text_features_to_speech("Hello", features, "output2.wav")
"""
def __init__(self,
model_path: str = 'Yue-Wang/BatonTTS-1.7B',
cosyvoice_model_dir: str = './pretrained_models/CosyVoice2-0.5B',
prompt_audio_path: str = './prompt.wav',
tensor_parallel_size: int = 1,
gpu_memory_utilization: float = 0.7,
fp16: bool = False):
"""
Initialize the Unified TTS Processor
Args:
model_path: Path to the main TTS model (Yue-Wang/BatonTTS-1.7B)
cosyvoice_model_dir: Directory path to CosyVoice2 model
prompt_audio_path: Path to prompt audio file for voice cloning
tensor_parallel_size: Number of GPUs for tensor parallelism
gpu_memory_utilization: GPU memory utilization ratio (0.0-1.0)
fp16: Whether to use half precision for CosyVoice2
Implementation:
1. Store configuration parameters
2. Initialize model loading flags
3. Defer actual model loading until first inference call
"""
self.model_path = model_path
self.cosyvoice_model_dir = cosyvoice_model_dir
self.prompt_audio_path = prompt_audio_path
self.tensor_parallel_size = tensor_parallel_size
self.gpu_memory_utilization = gpu_memory_utilization
self.fp16 = fp16
# Model components (loaded lazily)
self.llm = None
self.tokenizer = None
self.cosyvoice = None
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.sample_rate = None
# Cached prompt features
self.prompt_token = None
self.prompt_feat = None
self.speaker_embedding = None
# Special tokens and sampling parameters
self.special_tokens = {}
self.sampling_params = None
# Model loading flag
self.models_loaded = False
logger.info(f"UnifiedTTS initialized with model: {model_path}")
def _load_models(self):
"""
Load all required models for TTS processing
Model Components:
1. vLLM: Large language model for text-to-speech token generation
- Uses BATONTTS-1.7B model
- Handles both Mode 1 (text→features+speech) and Mode 2 (text+features→speech)
- Optimized with tensor parallelism for multi-GPU setups
2. Tokenizer: Text tokenization and encoding/decoding
- Handles special tokens for TTS control (<custom_token_0>, <custom_token_1>, etc.)
- Converts between text and token IDs for model input/output
- Supports offset correction for speech token recovery
3. CosyVoice2: Speech token to waveform conversion
- Converts discrete speech tokens to continuous audio waveforms
- Supports voice cloning using prompt audio features
- Configurable with half precision (fp16) for memory efficiency
Performance Optimizations:
- Tensor parallelism for multi-GPU inference
- GPU memory utilization control (default 0.7)
- Half precision support for CosyVoice2
- Lazy loading pattern to reduce initialization time
Usage:
This method is called automatically on first inference request.
All components are initialized once and reused for subsequent calls.
"""
if self.models_loaded:
return
logger.info("Loading models...")
# Load main TTS model with vLLM
# vLLM provides optimized inference for large language models with features like:
# - Continuous batching for improved throughput
# - PagedAttention for memory efficiency
# - Tensor parallelism for multi-GPU scaling
logger.info(f"Loading main model from {self.model_path}")
self.llm = LLM(
model=self.model_path,
tensor_parallel_size=self.tensor_parallel_size,
gpu_memory_utilization=self.gpu_memory_utilization,
trust_remote_code=True,
dtype="bfloat16",
max_model_len=2500,
)
# Load tokenizer separately
# The tokenizer handles conversion between text and token IDs
# It's essential for preparing model inputs and processing outputs
logger.info("Loading tokenizer...")
self.tokenizer = AutoTokenizer.from_pretrained(self.model_path)
# Load CosyVoice2 model
# CosyVoice2 converts discrete speech tokens to continuous audio waveforms
# It supports voice cloning using prompt audio features for consistent voice characteristics
logger.info(f"Loading CosyVoice2 model from {self.cosyvoice_model_dir}")
self.cosyvoice = CosyVoice2(cosyvoice_cache_dir, fp16=self.fp16, device="cuda" if torch.cuda.is_available() else "cpu")
self.sample_rate = self.cosyvoice.sample_rate
# Preload prompt audio features for voice cloning
# This extracts and caches features from the prompt audio file
# These features are used to maintain consistent voice characteristics
self._preload_prompt_features()
# Configure special tokens used for TTS control
# Special tokens structure the input/output format for different TTS modes
self._setup_special_tokens()
# Configure sampling parameters for text generation
# These parameters control the quality and diversity of generated speech tokens
self._setup_sampling_params()
self.models_loaded = True
logger.info("All models loaded successfully!")
def _preload_prompt_features(self):
"""
Preload prompt audio features for voice cloning
Feature Extraction Process:
1. Load prompt audio file and resample to 16kHz
2. Extract speech tokens using CosyVoice2 frontend
3. Resample audio to model's native sample rate
4. Extract speech features
5. Extract speaker embedding for voice characteristics
These features are cached and reused for all inference calls
to maintain consistent voice characteristics.
"""
if os.path.exists(self.prompt_audio_path):
try:
# Load and process prompt audio
prompt_speech = load_wav(self.prompt_audio_path, 16000)
# Extract speech tokens
self.prompt_token, _ = self.cosyvoice.frontend._extract_speech_token(prompt_speech)
logger.info(f"Preloaded prompt token, shape: {self.prompt_token.shape}")
# Extract speech features
prompt_speech_resample = torchaudio.transforms.Resample(
orig_freq=16000, new_freq=self.sample_rate
)(prompt_speech)
self.prompt_feat, _ = self.cosyvoice.frontend._extract_speech_feat(prompt_speech_resample)
logger.info(f"Preloaded prompt feat, shape: {self.prompt_feat.shape}")
# Extract speaker embedding
self.speaker_embedding = self.cosyvoice.frontend._extract_spk_embedding(prompt_speech)
logger.info(f"Preloaded speaker embedding, shape: {self.speaker_embedding.shape}")
except Exception as e:
logger.warning(f"Failed to load prompt audio: {e}")
self._use_default_prompt_features()
else:
logger.warning(f"Prompt audio not found: {self.prompt_audio_path}")
self._use_default_prompt_features()
def _use_default_prompt_features(self):
"""
Use default (empty) prompt features when prompt audio is unavailable
Default Features:
- Empty speech token tensor
- Zero-filled feature tensor with correct dimensions
- Zero-filled speaker embedding with standard size
These defaults allow the model to generate speech without
voice cloning, using the model's default voice characteristics.
"""
self.prompt_token = torch.zeros(1, 0, dtype=torch.int32)
self.prompt_feat = torch.zeros(1, 0, 80)
self.speaker_embedding = torch.zeros(1, 192)
logger.info("Using default prompt features")
def _setup_special_tokens(self):
"""
Configure special tokens for TTS control and mode switching
Special Tokens Purpose:
- custom_token_0: Marks the start of speech token generation phase
* In Mode 1: Signals transition from text processing to speech token output
* In Mode 2: Indicates where speech tokens should be inserted
- custom_token_1: Marks the end of speech token generation phase
* Used as stop token to terminate speech token generation
* Prevents model from generating beyond intended speech sequence
- custom_token_2: Additional control token for future extensions
* Reserved for advanced TTS control features
* Currently used in Mode 2 for feature conditioning
- eos_token_id: Standard end-of-sequence marker
* Separates different parts of the input sequence
* Used in prompt construction for both modes
Token Format Compatibility:
- Stores tokens as lists to match mode1.py format exactly
- Maintains backward compatibility with dictionary format
- Ensures consistent token handling across different inference modes
Implementation Details:
- Tokens are encoded using tokenizer's call method (matching mode1.py)
- Each token is stored as a list of token IDs (usually single element)
- Dictionary format provides easy access for legacy code
"""
# Store as lists (matching mode1.py format)
# This ensures exact compatibility with the original mode1.py implementation
self.custom_token_0_ids = self.tokenizer('<custom_token_0>').input_ids
self.custom_token_1_ids = self.tokenizer('<custom_token_1>').input_ids
self.custom_token_2_ids = self.tokenizer('<custom_token_2>').input_ids
self.eos_token_id = self.tokenizer.eos_token_id
# Maintain dictionary format for backward compatibility
# This allows existing code to continue using self.special_tokens['token_name']
self.special_tokens = {
'custom_token_0': self.custom_token_0_ids[0],
'custom_token_1': self.custom_token_1_ids[0],
'custom_token_2': self.custom_token_2_ids[0],
'eos_token_id': self.eos_token_id
}
logger.info("Special tokens configured")
def _setup_sampling_params(self):
"""
Configure sampling parameters for text generation (matches mode1.py exactly)
Sampling Strategy:
The parameters are carefully tuned for TTS generation to balance quality,
diversity, and computational efficiency. These settings match mode1.py
to ensure consistent behavior across different inference implementations.
Parameter Details:
- temperature (0.8): Controls randomness in token selection
* Lower values (0.1-0.5): More deterministic, consistent output
* Higher values (0.8-1.0): More creative, diverse output
* 0.6 provides optimal balance for natural speech variation
- top_p (1.0): Nucleus sampling threshold for quality control
* Considers all tokens in the probability distribution
* Allows full model expressiveness for speech token generation
* Maintains coherent speech token sequences
- max_tokens (2048): Maximum tokens to generate in one pass
* Accommodates long speech sequences (up to ~20 seconds of audio)
* Matches mode1.py configuration for consistency
* Prevents runaway generation while allowing sufficient length
- stop_token_ids: Tokens that terminate generation
* Uses custom_token_1 to mark end of speech token sequence
* Ensures clean separation between speech tokens and other content
* Prevents model from generating beyond intended speech boundary
- repetition_penalty (1.1): Reduces repetitive output patterns
* Slightly penalizes recently generated tokens
* Improves speech naturalness by reducing monotonous patterns
* Maintains speech quality while encouraging variation
Usage:
These parameters are used by vLLM during text generation for both
Mode 1 (text→speech tokens) and Mode 2 (text+features→speech tokens).
"""
self.sampling_params = SamplingParams(
temperature=0.6, # Balanced randomness for natural speech variation
top_p=1, # Full nucleus sampling for maximum expressiveness
max_tokens=2048, # Maximum sequence length (matches mode1.py)
stop_token_ids=[self.eos_token_id], # Stop at end of speech tokens
repetition_penalty=1.1, # Slight penalty to reduce repetitive patterns
)
logger.info("Sampling parameters configured")
def _prepare_mode1_input(self, text: str) -> str:
"""
Prepare input prompt for Mode 1 inference (matches mode1.py exactly)
Mode 1 Workflow:
Mode 1 performs end-to-end text-to-speech conversion in a single pass.
The model generates both intermediate features and final speech tokens
from the input text, making it suitable for scenarios where no specific
voice characteristics are required.
Input Format Structure: <custom_token_0> + text + <custom_token_1>
Component Breakdown:
1. <custom_token_0>: Speech generation trigger
- Signals model to begin speech token generation
- Marks the start of the input sequence
- Model generates speech tokens after this marker
2. text: The input text to be converted to speech
- Can be any natural language text
- Processed by the model's text understanding capabilities
- No length restrictions beyond model's context window
3. <custom_token_1>: End-of-text marker
- Signals the end of text input to the model
- Separates text content from control tokens
- Essential for proper model interpretation
Processing Steps:
1. Tokenize input text using the model's tokenizer
2. Construct sequence: custom_token_0 + text_tokens + custom_token_1
3. Decode token sequence back to text for vLLM processing
Compatibility:
This implementation exactly matches mode1.py's prepare_inference_prompt
method to ensure identical behavior and output quality.
Args:
text: Input text to synthesize
Returns:
Formatted prompt string for vLLM inference
"""
# Encode text to tokens (matching mode1.py implementation)
# Using return_tensors="pt" and converting to list for consistency
text_tokens = self.tokenizer(text, return_tensors="pt").input_ids[0].tolist()
# Construct input sequence: <custom_token_0> + text + <custom_token_1>
# This exactly matches mode1.py's prepare_inference_prompt method
# The sequence structure is critical for proper model interpretation
input_sequence = (
self.custom_token_0_ids + # Speech generation trigger
text_tokens + # Original text content
self.custom_token_1_ids # End-of-text marker
)
# Convert token sequence back to text for vLLM processing
# vLLM expects text input, so we decode the token sequence
prompt = self.tokenizer.decode(input_sequence, skip_special_tokens=False)
return prompt
def _prepare_mode2_input(self, text: str, word_features: str) -> str:
"""
Prepare input prompt for Mode 2 inference (matches mode2.py exactly)
Mode 2 Input Format (from mode2.py):
<custom_token_0> + text + <custom_token_1> + generated_features + <custom_token_2>
Args:
text: Input text to synthesize
word_features: JSON string containing word-level prosodic features (called generated_features in mode2.py)
Returns:
Formatted prompt string for vLLM inference
Implementation Logic:
1. Tokenize input text and generated_features using the model's tokenizer
2. Construct sequence: custom_token_0 + text_tokens + custom_token_1 + features_tokens + custom_token_2
3. Decode back to text format for vLLM processing
4. Return formatted prompt
The model will generate speech_tokens ending with eos_token_id
This format matches the exact implementation in mode2.py
"""
# Encode text and generated_features to tokens (matching mode2.py implementation)
text_tokens = self.tokenizer(text, return_tensors="pt").input_ids[0].tolist()
features_tokens = self.tokenizer(word_features, return_tensors="pt").input_ids[0].tolist()
# Construct input sequence: custom_token_0 + text + custom_token_1 + generated_features + custom_token_2
# This exactly matches mode2.py's prepare_mode2_inference_prompt method
input_sequence = (
[self.special_tokens['custom_token_0']] +
text_tokens +
[self.special_tokens['custom_token_1']] +
features_tokens +
[self.special_tokens['custom_token_2']]
)
# Convert token sequence back to text for vLLM
prompt = self.tokenizer.decode(input_sequence, skip_special_tokens=False)
return prompt
def _process_features(self, features: str) -> str:
"""
Process word features input (JSON parsing with fallback)
Processing Logic:
1. Attempt to parse as JSON for validation
2. If successful, convert back to string representation
3. If parsing fails, use original string as-is
4. Log the processing result for debugging
Args:
features: Input features string (JSON or plain text)
Returns:
Processed features string
This flexible approach handles both JSON-formatted features
and plain text features, ensuring compatibility with various
input formats while maintaining robustness.
"""
try:
# Try JSON parsing for validation
parsed_features = json.loads(features)
processed_features = str(parsed_features)
logger.debug("Successfully parsed features as JSON")
return processed_features
except (json.JSONDecodeError, TypeError, ValueError) as e:
# Use original string if JSON parsing fails
logger.debug(f"Using features as plain string: {e}")
return features
def _generate_response(self, prompt: str) -> Tuple[str, bool]:
"""
Generate model response using vLLM
Generation Process:
1. Use vLLM's generate method with configured sampling parameters
2. Extract generated text from model output
3. Check if generation was truncated due to length limits
4. Return both generated text and truncation status
Args:
prompt: Formatted input prompt for the model
Returns:
Tuple of (generated_text, is_truncated)
The truncation flag helps identify potential quality issues
when the generation hits the maximum token limit.
"""
outputs = self.llm.generate([prompt], self.sampling_params)
# Extract generated text and check truncation
output = outputs[0]
generated_text = output.outputs[0].text
is_truncated = output.outputs[0].finish_reason == 'length'
if is_truncated:
logger.warning("Generation was truncated, may affect audio quality")
return generated_text, is_truncated
def _extract_mode1_features(self, generated_text: str, original_prompt: str) -> str:
"""
Extract generated features from Mode 1 output
Mode 1 Output Format:
According to the sequence format: custom_token_0 + text + custom_token_1 + generated_features + custom_token_2 + features + eos_token_id
This method extracts the generated_features part from the model output.
The generated_features are located between custom_token_1 and custom_token_2.
Args:
generated_text: Model generated text
original_prompt: Original input prompt
Returns:
JSON string containing the generated features
"""
try:
# Find the features section between custom_token_1 and custom_token_2
custom_token_1_str = self.tokenizer.decode(self.custom_token_1_ids, skip_special_tokens=False)
custom_token_2_str = self.tokenizer.decode(self.custom_token_2_ids, skip_special_tokens=False)
# Look for features in the generated text
full_text = original_prompt + generated_text
# Find the start of features (after custom_token_1)
start_idx = full_text.find(custom_token_1_str)
if start_idx == -1:
logger.warning("Could not find custom_token_1 in generated text")
return ""
start_idx += len(custom_token_1_str)
# Find the end of features (before custom_token_2)
end_idx = full_text.find(custom_token_2_str, start_idx)
if end_idx == -1:
logger.warning("Could not find custom_token_2 in generated text")
# If no custom_token_2 found, try to find eos_token
eos_str = self.tokenizer.decode([self.eos_token_id], skip_special_tokens=False)
end_idx = full_text.find(eos_str, start_idx)
if end_idx == -1:
# Take the rest of the text
end_idx = len(full_text)
# Extract features text
features_text = full_text[start_idx:end_idx].strip()
if features_text:
logger.info(f"Extracted features: {features_text[:100]}...")
return features_text
else:
logger.warning("No features found in generated text")
return ""
except Exception as e:
logger.error(f"Error extracting mode1 features: {e}")
return ""
def _extract_mode1_outputs(self, generated_text: str, original_prompt: str) -> List[int]:
"""
Extract speech tokens from Mode 1 generation output (matches mode1.py exactly)
Output Processing Workflow:
Mode 1 generates a sequence containing both the original prompt and new speech tokens.
This method extracts only the speech tokens that represent the audio content,
filtering out text tokens and control tokens to prepare for audio synthesis.
Token Offset Mechanism:
The model uses a token offset system to distinguish between different token types:
- Text tokens: IDs in lower range (0 to ~151669)
- Speech tokens: IDs in higher range (151669+ with offset applied)
- The offset (151669 + 100) is used to map speech tokens back to their original values
Processing Steps:
1. Merge prompt and generated text to get complete sequence
- Prompt contains: text + [EOS] + <custom_token_0>
- Generated text contains: speech_tokens + <custom_token_1>
- Full sequence represents the complete model output
2. Encode merged text to token IDs
- Converts text representation back to numerical token IDs
- Preserves all tokens including special tokens and speech tokens
- Uses same tokenizer as model for consistency
3. Apply offset correction to recover speech tokens
- Speech tokens are stored with offset (151669 + 100) added
- Subtracting offset recovers original speech token values
- Only tokens above offset threshold are considered speech tokens
4. Filter and return speech tokens
- Removes text tokens, control tokens, and invalid tokens
- Returns clean list of speech token IDs for audio synthesis
Compatibility:
This implementation exactly matches mode1.py's extract_output_audio_tokens
method to ensure identical behavior, output format, and audio quality.
Args:
generated_text: Model generated text
original_prompt: Original input prompt
Returns:
List of speech token integers (offset corrected)
"""
try:
# Step 1: Merge original prompt and generated text (matching mode1.py exactly)
# This gives us the complete sequence that the model processed
full_text = original_prompt + generated_text
if not full_text.strip():
return []
# Step 2: Encode text to token IDs (matching mode1.py tokenization)
# Convert the text back to numerical token IDs for processing
token_ids = self.tokenizer.encode(full_text, add_special_tokens=False)
# Step 3: Apply offset correction to recover original speech tokens
# The model stores speech tokens with an offset of (151669 + 100)
# We subtract this offset to get the original speech token values
speech_tokens = []
offset = 151669 + 100 # Exact same offset as mode1.py
for token_id in token_ids:
if token_id >= offset:
# Recover original speech token by subtracting offset
original_token = token_id - offset
speech_tokens.append(original_token)
logger.info(f"Extracted speech_tokens count: {len(speech_tokens)}")
return speech_tokens
except Exception as e:
logger.error(f"Error extracting mode1 outputs: {e}")
return []
def _extract_mode2_outputs(self, generated_text: str, original_prompt: str) -> List[int]:
"""
Extract speech_tokens from Mode 2 output (matches mode2.py exactly)
Extraction Logic (from mode2.py extract_speech_tokens_from_mode2_output):
1. Encode generated text to token IDs using tokenizer
2. Apply offset correction to recover original speech token values
3. Filter tokens to ensure they're in valid speech token range
Args:
generated_text: Model generated text
original_prompt: Original input prompt (unused in mode2.py, kept for consistency)
Returns:
List of speech token integers (offset corrected)
This exactly matches the implementation in mode2.py's extract_speech_tokens_from_mode2_output method
"""
try:
if not generated_text.strip():
return []
# Encode generated text to token IDs (matching mode2.py)
token_ids = self.tokenizer.encode(generated_text, add_special_tokens=False)
# Recover original speech_token values (subtract offset)
speech_tokens = []
offset = 151669 + 100 # Consistent with mode2.py implementation
for token_id in token_ids:
if token_id >= offset:
original_token = token_id - offset
speech_tokens.append(original_token)
logger.info(f"Extracted speech_tokens count: {len(speech_tokens)}")
return speech_tokens
except Exception as e:
logger.error(f"Error extracting mode2 outputs: {e}")
return []
def _text_to_speech_tokens(self, text: str) -> List[int]:
"""
Convert text representation to speech token integers
Conversion Process:
1. Tokenize text using the model's tokenizer
2. Apply offset correction to recover original token values
3. Filter tokens to ensure they're in valid speech token range
4. Return list of corrected integer tokens
Args:
text: Text representation of speech tokens
Returns:
List of speech token integers
The offset correction (151669 + 100) reverses the encoding
applied during model training to map speech tokens to the
model's vocabulary space.
"""
if not text.strip():
return []
try:
# Encode text to token IDs
token_ids = self.tokenizer.encode(text, add_special_tokens=False)
# Apply offset correction to recover original speech tokens
speech_tokens = []
offset = 151669 + 100 # Consistent with original implementation
for token_id in token_ids:
if token_id >= offset:
original_token = token_id - offset
speech_tokens.append(original_token)
return speech_tokens
except Exception as e:
logger.error(f"Text to speech_token conversion failed: {e}")
return []
def _convert_tokens_to_audio(self, speech_tokens: List[int], speed: float = 1.0) -> Optional[torch.Tensor]:
"""
Convert speech tokens to audio waveform
Conversion Process:
1. Validate speech tokens and convert to tensor format
2. Generate unique inference ID for cache management
3. Call CosyVoice2's token2wav method with prompt features
4. Clean up inference cache to prevent memory leaks
5. Return generated audio tensor
Args:
speech_tokens: List of speech token integers
speed: Speech speed multiplier (1.0 = normal speed)
Returns:
Audio tensor or None if conversion fails
The cache management ensures proper cleanup of temporary
data structures used during audio generation.
"""
if len(speech_tokens) == 0:
logger.warning("Empty speech tokens provided")
return None
try:
# Convert to tensor format
speech_token_tensor = torch.tensor([speech_tokens], dtype=torch.int32)
logger.debug(f"Converting speech tokens, shape: {speech_token_tensor.shape}")
# Generate unique inference ID
inference_uuid = str(uuid.uuid1())
# Initialize cache
with self.cosyvoice.model.lock:
self.cosyvoice.model.hift_cache_dict[inference_uuid] = None
try:
# Generate audio using CosyVoice2
tts_speech = self.cosyvoice.model.token2wav(
token=speech_token_tensor,
prompt_token=self.prompt_token,
prompt_feat=self.prompt_feat,
embedding=self.speaker_embedding,
token_offset=0,
uuid=inference_uuid,
finalize=True,
speed=speed
)
return tts_speech.cpu()
finally:
# Clean up cache
with self.cosyvoice.model.lock:
if inference_uuid in self.cosyvoice.model.hift_cache_dict:
self.cosyvoice.model.hift_cache_dict.pop(inference_uuid)
except Exception as e:
logger.error(f"Audio conversion failed: {e}")
return None
def _save_audio(self, audio_tensor: torch.Tensor, output_path: str) -> bool:
"""
Save audio tensor to file
Saving Process:
1. Ensure output directory exists
2. Use torchaudio to save tensor as WAV file
3. Apply correct sample rate from CosyVoice2 model
4. Handle any file I/O errors gracefully
Args:
audio_tensor: Generated audio waveform tensor
output_path: Target file path for saving
Returns:
True if saving successful, False otherwise
The method ensures proper directory creation and error handling
for robust file operations across different environments.
"""
# Ensure output directory exists
try:
os.makedirs(os.path.dirname(output_path), exist_ok=True)
except Exception:
pass
try:
# Save audio file
torchaudio.save(output_path, audio_tensor, self.sample_rate)
logger.info(f"Audio saved successfully to: {output_path}")
return True
except Exception as e:
logger.error(f"Failed to save audio: {e}")
return False
def text_to_speech(self, text: str, output_path: str, speed: float = 1.0) -> bool:
"""
Mode 1: End-to-End Text-to-Speech Conversion (matches mode1.py exactly)
Overview:
Mode 1 provides a complete text-to-speech pipeline that generates speech directly
from input text without requiring external features or prompt audio. The model
internally generates both acoustic features and speech tokens in a single pass,
making it ideal for general-purpose TTS applications.
Key Advantages:
- Single-pass inference for optimal performance
- No dependency on external features or prompt audio
- Consistent voice characteristics across different texts
- Simplified API for straightforward TTS tasks
- Direct compatibility with mode1.py implementation
Technical Workflow:
1. Model Initialization: Ensures all components are loaded and ready
- vLLM engine for language model inference
- Tokenizer for text processing and token conversion
- CosyVoice2 for speech synthesis from tokens
2. Input Validation and Preparation: Formats text according to Mode 1 protocol
- Validates input text for proper content
- Constructs prompt: text + [EOS] + <custom_token_0>
- This format signals the model to generate speech tokens directly
3. Speech Token Generation: Uses vLLM for efficient model inference
- Applies optimized sampling parameters for quality control
- Generates speech token sequence ending with <custom_token_1>
- Produces discrete tokens representing audio content
4. Token Extraction and Processing: Isolates speech tokens from output
- Filters out input text and control tokens
- Applies offset correction (151669 + 100) to recover original values
- Validates token sequence for successful audio synthesis
5. Audio Synthesis: Converts tokens to waveform using CosyVoice2
- Transforms discrete speech tokens to continuous audio signal
- Applies speed adjustment if specified (default 1.0 = normal)
- Generates high-quality audio with natural voice characteristics
6. File Output: Saves audio to specified path with proper formatting
- Supports common audio formats (WAV, MP3, etc.)
- Maintains audio quality and sample rate consistency
- Provides success/failure feedback for error handling
Error Handling and Validation:
- Comprehensive input validation for text content
- Model loading failure detection and recovery
- Token generation validation and error reporting
- Audio synthesis error handling with detailed logging
- File I/O error management with clear feedback
Performance Optimizations:
- Lazy model loading to reduce initialization overhead
- Efficient token processing to minimize memory usage
- Optimized sampling parameters for quality-speed balance
- Streamlined audio conversion pipeline
Compatibility:
This implementation exactly matches mode1.py's behavior to ensure:
- Identical input format and processing logic
- Same token extraction and offset correction
- Consistent audio quality and characteristics
- Compatible output format and file structure
Args:
text: Input text to convert to speech
- Supports natural language text in multiple languages
- Handles punctuation, numbers, and special characters
- No strict length limitations (within model context window)
- Automatically processes formatting and normalization
output_path: File path where generated audio will be saved
- Supports various audio formats (WAV, MP3, FLAC, etc.)
- Creates parent directories if they don't exist
- Overwrites existing files at the specified path
- Should include appropriate file extension
speed: Speech speed multiplier for tempo control
- 1.0 = normal speaking speed (default)
- < 1.0 = slower speech (e.g., 0.8 for 20% slower)
- > 1.0 = faster speech (e.g., 1.2 for 20% faster)
- Applied during audio synthesis phase
Returns:
bool: Success status of the TTS operation
- True: Audio successfully generated and saved
- False: Operation failed (check logs for details)
Raises:
Exception: Caught and logged, returns False for graceful handling
Usage Examples:
# Basic text-to-speech
success = tts.text_to_speech("Hello world", "output.wav")
# With custom speed
success = tts.text_to_speech(
"Welcome to our service",
"welcome.wav",
speed=0.9
)
"""
try:
# Ensure models are loaded
self._load_models()
logger.info(f"Mode 1: Text to speech - {text}")
# Validate input
if not text.strip():
logger.error("Empty text provided")
return False
# Prepare input for Mode 1
prompt = self._prepare_mode1_input(text)
# Generate response
generated_text, is_truncated = self._generate_response(prompt)
# Extract speech_tokens (Mode 1 now only returns speech_tokens, matching mode1.py)
speech_tokens = self._extract_mode1_outputs(generated_text, prompt)
if len(speech_tokens) == 0:
logger.error("Failed to extract speech tokens")
return False
# Convert to audio
audio_tensor = self._convert_tokens_to_audio(speech_tokens, speed)
if audio_tensor is None:
logger.error("Failed to convert tokens to audio")
return False
# Save audio file
success = self._save_audio(audio_tensor, output_path)
if success:
logger.info(f"Mode 1 synthesis completed successfully!")
return True
else:
logger.error("Failed to save audio file")
return False
except Exception as e:
logger.error(f"Mode 1 synthesis failed: {e}")
return False
def text_to_speech_with_features(self, text: str, output_path: str, speed: float = 1.0) -> Tuple[bool, str]:
"""
Mode 1: End-to-End Text-to-Speech Conversion with Features Extraction
This method extends the standard text_to_speech functionality to also return
the generated features that the model produces internally. This is useful
for understanding the prosodic characteristics the model assigns to the text.
Args:
text: Input text to convert to speech
output_path: File path where generated audio will be saved
speed: Speech speed multiplier for tempo control
Returns:
Tuple[bool, str]: (success_status, generated_features_json)
- success_status: True if audio generation succeeded, False otherwise
- generated_features_json: JSON string of generated features, empty if failed
"""
try:
# Ensure models are loaded
self._load_models()
logger.info(f"Mode 1 with features: Text to speech - {text}")
# Validate input
if not text.strip():
logger.error("Empty text provided")
return False, ""
# Prepare input for Mode 1
prompt = self._prepare_mode1_input(text)
# Generate response
generated_text, is_truncated = self._generate_response(prompt)
# Extract features from the generated text
generated_features = self._extract_mode1_features(generated_text, prompt)
# Extract speech_tokens (Mode 1 now only returns speech_tokens, matching mode1.py)
speech_tokens = self._extract_mode1_outputs(generated_text, prompt)
if len(speech_tokens) == 0:
logger.error("Failed to extract speech tokens")
return False, generated_features
# Convert to audio
audio_tensor = self._convert_tokens_to_audio(speech_tokens, speed)
if audio_tensor is None:
logger.error("Failed to convert tokens to audio")
return False, generated_features
# Save audio file
success = self._save_audio(audio_tensor, output_path)
if success:
logger.info(f"Mode 1 synthesis with features completed successfully!")
return True, generated_features
else:
logger.error("Failed to save audio file")
return False, generated_features
except Exception as e:
logger.error(f"Mode 1 synthesis with features failed: {e}")
return False, ""
def text_features_to_speech(self, text: str, word_features: str, output_path: str,
speed: float = 1.0) -> bool:
"""
Mode 2: Convert text + features to speech (uses provided features)
Process Flow:
1. Load models if not already loaded
2. Prepare Mode 2 input format with text and features
3. Generate model response to get speech_tokens only
4. Extract speech_tokens from the generated output
5. Convert speech_tokens to audio waveform
6. Save audio to specified output path
Args:
text: Input text to synthesize
word_features: Pre-generated word-level prosodic features
output_path: Path to save the generated audio file
speed: Speech speed multiplier (1.0 = normal)
Returns:
True if synthesis successful, False otherwise
This mode is ideal when you have specific prosodic requirements
and want precise control over speech characteristics.
"""
try:
# Ensure models are loaded
self._load_models()
logger.info(f"Mode 2: Text + features to speech - {text}")
logger.info(f"Features: {word_features[:100]}...") # Show first 100 chars
# Validate inputs
if not text.strip() or not word_features.strip():
logger.error("Empty text or features provided")
return False
# Prepare input for Mode 2
prompt = self._prepare_mode2_input(text, word_features)
# Generate response
generated_text, is_truncated = self._generate_response(prompt)
# Extract speech_tokens (Mode 2 now matches mode2.py implementation)
speech_tokens = self._extract_mode2_outputs(generated_text, prompt)
if len(speech_tokens) == 0:
logger.error("Failed to extract speech tokens")
return False
# Convert to audio
audio_tensor = self._convert_tokens_to_audio(speech_tokens, speed)
if audio_tensor is None:
logger.error("Failed to convert tokens to audio")
return False
# Save audio file
success = self._save_audio(audio_tensor, output_path)
if success:
logger.info(f"Mode 2 synthesis completed successfully!")
return True
else:
logger.error("Failed to save audio file")
return False
except Exception as e:
logger.error(f"Mode 2 synthesis failed: {e}")
return False
def cleanup(self):
"""
Clean up model resources and free memory
Cleanup Process:
1. Delete vLLM model instance
2. Delete CosyVoice2 model instance
3. Clear CUDA cache if available
4. Reset loading flags and cached features
This method should be called when the TTS instance is no longer
needed to free up GPU memory for other processes.
"""
if self.llm:
del self.llm
self.llm = None
if self.cosyvoice:
del self.cosyvoice
self.cosyvoice = None
if torch.cuda.is_available():
torch.cuda.empty_cache()
self.models_loaded = False
logger.info("Model resources cleaned up")
def main():
"""
Command-line interface for testing both TTS modes
CLI Features:
1. Supports both Mode 1 (text-only) and Mode 2 (text+features)
2. Uses default configurations from single_inference_tts.py
3. Provides comprehensive argument parsing for all parameters
4. Includes example test cases for both modes
Usage Examples:
# Mode 1: Text only
python unified_tts.py --mode 1 --text "Hello world" --output "output1.wav"
# Mode 2: Text + features
python unified_tts.py --mode 2 --text "Hello world" --features '[{...}]' --output "output2.wav"
The CLI provides a convenient way to test the TTS functionality
and serves as an example for integration into other scripts.
"""
parser = argparse.ArgumentParser(description='Unified TTS Script - Two modes for text-to-speech synthesis')
# Model configuration arguments
parser.add_argument('--model_path', type=str, default='Yue-Wang/BatonTTS-1.7B',
help='Path to the main TTS model')
parser.add_argument('--cosyvoice_model_dir', type=str, default='./pretrained_models/CosyVoice2-0.5B',
help='Directory path to CosyVoice2 model')
parser.add_argument('--prompt_audio_path', type=str, default='./prompt.wav',
help='Path to prompt audio file for voice cloning')
# Mode selection and input arguments
parser.add_argument('--mode', type=int, choices=[1, 2], default=1,
help='TTS mode: 1=text only, 2=text+features')
parser.add_argument('--text', type=str, default='Kids are talking by the door',
help='Input text to synthesize')
parser.add_argument('--features', type=str,
default='[{"word": "Kids are talking","pitch_mean": 315,"pitch_slope": 90,"energy_rms": 0.005,"energy_slope": 25,"spectral_centroid": 2650},{"word": "by the door","pitch_mean": 360,"pitch_slope": -110,"energy_rms": 0.004,"energy_slope": -30,"spectral_centroid": 2900}]',
help='Word-level features for Mode 2 (JSON format)')
# Output and performance arguments
parser.add_argument('--output', type=str, default='unified_output.wav',
help='Output audio file path')
parser.add_argument('--speed', type=float, default=1.0,
help='Speech speed multiplier (1.0 = normal speed)')
parser.add_argument('--tensor_parallel_size', type=int, default=1,
help='Number of GPUs for tensor parallelism')
parser.add_argument('--gpu_memory_utilization', type=float, default=0.7,
help='GPU memory utilization ratio (0.0-1.0)')
parser.add_argument('--fp16', action='store_true',
help='Use half precision for CosyVoice2')
args = parser.parse_args()
# Initialize TTS processor
tts = UnifiedTTS(
model_path=args.model_path,
cosyvoice_model_dir=args.cosyvoice_model_dir,
prompt_audio_path=args.prompt_audio_path,
tensor_parallel_size=args.tensor_parallel_size,
gpu_memory_utilization=args.gpu_memory_utilization,
fp16=args.fp16
)
try:
# Execute synthesis based on selected mode
if args.mode == 1:
print(f"Running Mode 1: Text to Speech")
print(f"Text: {args.text}")
print(f"Output: {args.output}")
success, features = tts.text_to_speech_with_features(
text=args.text,
output_path=args.output,
speed=args.speed
)
if success:
print(f"Features: {features}")
elif args.mode == 2:
print(f"Running Mode 2: Text + Features to Speech")
print(f"Text: {args.text}")
print(f"Features: {args.features[:100]}...") # Show first 100 chars
print(f"Output: {args.output}")
success = tts.text_features_to_speech(
text=args.text,
word_features=args.features,
output_path=args.output,
speed=args.speed
)
# Report results
if success:
print(f"\n✅ Synthesis completed successfully!")
print(f"Audio file saved to: {args.output}")
else:
print(f"\n❌ Synthesis failed!")
print(f"Please check the logs for error details.")
sys.exit(1)
except KeyboardInterrupt:
print("\n⏹️ Synthesis interrupted by user")
except Exception as e:
print(f"\n💥 Unexpected error: {e}")
sys.exit(1)
finally:
# Clean up resources
tts.cleanup()
print("🧹 Resources cleaned up")
if __name__ == "__main__":
main()
|