File size: 22,539 Bytes
b48d7b7 |
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 |
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ β DETAILED SOURCE FILE LISTING BY CATEGORY β ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ MAIN INFERENCE PIPELINE FILES βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ /home/user/IndexTTS-Rust/indextts/infer_v2.py (739 LINES) βββ CRITICAL ββ Purpose: Main TTS inference class (IndexTTS2) ββ Key Classes: β ββ QwenEmotion (emotion text-to-vector conversion) β ββ IndexTTS2 (main inference class) β ββ Helper functions for emotion/audio processing ββ Key Methods: β ββ __init__() - Initialize all models and codecs β ββ infer() - Single text generation with emotion control β ββ infer_fast() - Parallel segment generation β ββ get_emb() - Extract semantic embeddings β ββ remove_long_silence() - Silence token removal β ββ insert_interval_silence() - Silence insertion β ββ Cache management for repeated generation ββ Models Loaded: β ββ UnifiedVoice (GPT model for mel token generation) β ββ W2V-BERT (semantic feature extraction) β ββ RepCodec (semantic codec) β ββ S2Mel model (semantic-to-mel conversion) β ββ CAMPPlus (speaker embedding) β ββ BigVGAN vocoder β ββ Qwen-based emotion model β ββ Emotion/speaker matrices ββ External Dependencies: torch, transformers, librosa, safetensors /home/user/IndexTTS-Rust/webui.py (18KB) βββ WEB INTERFACE ββ Purpose: Gradio-based web UI for IndexTTS ββ Key Components: β ββ Model initialization (IndexTTS2 instance) β ββ Language selection (Chinese/English) β ββ Emotion control modes (4 modes) β ββ Example case loading from cases.jsonl β ββ Progress bar integration β ββ Output management ββ Features: β ββ Real-time inference β ββ Multiple emotion control methods β ββ Batch processing β ββ Task caching β ββ i18n support β ββ Pre-loaded example cases ββ Web Framework: Gradio 5.34.1 /home/user/IndexTTS-Rust/indextts/cli.py (64 LINES) ββ Purpose: Command-line interface ββ Usage: python -m indextts.cli <text> -v <voice.wav> -o <output.wav> [options] ββ Arguments: β ββ text: Text to synthesize β ββ -v/--voice: Voice reference audio β ββ -o/--output_path: Output file path β ββ -c/--config: Config file path β ββ --model_dir: Model directory β ββ --fp16: Use FP16 precision β ββ -d/--device: Device (cpu/cuda/mps/xpu) β ββ -f/--force: Force overwrite ββ Uses: IndexTTS (v1 model) TEXT PROCESSING & NORMALIZATION FILES βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ /home/user/IndexTTS-Rust/indextts/utils/front.py (700 LINES) βββ CRITICAL ββ Purpose: Text normalization and tokenization ββ Key Classes: β ββ TextNormalizer (700+ lines) β β ββ Pattern Definitions: β β β ββ PINYIN_TONE_PATTERN (regex for pinyin with tones 1-5) β β β ββ NAME_PATTERN (regex for Chinese names) β β β ββ ENGLISH_CONTRACTION_PATTERN (regex for 's contractions) β β ββ Methods: β β β ββ normalize() - Main normalization β β β ββ use_chinese() - Language detection β β β ββ save_pinyin_tones() - Extract pinyin with tones β β β ββ restore_pinyin_tones() - Restore pinyin β β β ββ save_names() - Extract names β β β ββ restore_names() - Restore names β β β ββ correct_pinyin() - Phoneme correction (jqxβv) β β β ββ char_rep_map - Character replacement dictionary β β ββ Normalizers: β β ββ zh_normalizer (Chinese) - Uses WeTextProcessing/wetext β β ββ en_normalizer (English) - Uses tn library β β β ββ TextTokenizer (200+ lines) β ββ Methods: β β ββ encode() - Text to token IDs β β ββ decode() - Token IDs to text β β ββ convert_tokens_to_ids() β β ββ convert_ids_to_tokens() β β ββ Vocab management β ββ Special Tokens: β β ββ BOS: "<s>" (ID 0) β β ββ EOS: "</s>" (ID 1) β β ββ UNK: "<unk>" β ββ Tokenizer: SentencePiece (BPE-based) ββ Language Support: β ββ Chinese (simplified & traditional) β ββ English β ββ Mixed Chinese-English ββ Critical Pattern Matching: ββ Pinyin tone detection ββ Name entity detection ββ Email matching ββ Character replacement ββ Punctuation handling GPT MODEL ARCHITECTURE FILES βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ /home/user/IndexTTS-Rust/indextts/gpt/model_v2.py (747 LINES) βββ CRITICAL ββ Purpose: UnifiedVoice GPT-based TTS model ββ Key Classes: β ββ UnifiedVoice (700+ lines) β β ββ Architecture: β β β ββ Input Embeddings: Text (256 vocab), Mel (8194 vocab) β β β ββ Position Embeddings: Learned embeddings for mel/text β β β ββ GPT Transformer: Configurable layers/heads β β β ββ Conditioning Encoder: Conformer or Perceiver-based β β β ββ Emotion Conditioning: Separate conformer + perceiver β β β ββ Output Heads: Text prediction, Mel prediction β β β β β ββ Parameters: β β β ββ layers: 8 (transformer depth) β β β ββ model_dim: 512 (embedding dimension) β β β ββ heads: 8 (attention heads) β β β ββ max_text_tokens: 120 β β β ββ max_mel_tokens: 250 β β β ββ number_mel_codes: 8194 β β β ββ condition_type: "conformer_perceiver" or "conformer_encoder" β β β ββ Various activation functions β β β β β ββ Key Methods: β β β ββ forward() - Forward pass β β β ββ post_init_gpt2_config() - Initialize for inference β β β ββ generate_mel() - Mel token generation β β β ββ forward_with_cond_scale() - With classifier-free guidance β β β ββ Cache management β β β β β ββ Conditioning System: β β ββ Speaker conditioning via mel spectrogram β β ββ Conformer encoder for speaker features β β ββ Perceiver for attention pooling β β ββ Emotion conditioning (separate pathway) β β ββ Emotion vector support (8-dimensional) β β β ββ ResBlock (40+ lines) β β ββ Conv1d layers with GroupNorm β β ββ ReLU activation with residual connection β β β ββ GPT2InferenceModel (200+ lines) β β ββ Inference wrapper for GPT2 β β ββ KV cache support β β ββ Model parallelism support β β ββ Token-by-token generation β β β ββ ConditioningEncoder (30 lines) β β ββ Conv1d initialization β β ββ Attention blocks β β ββ Optional mean pooling β β β ββ MelEncoder (30 lines) β β ββ Conv1d layers β β ββ ResBlocks β β ββ 4x reduction β β β ββ LearnedPositionEmbeddings (15 lines) β β ββ Learnable positional embeddings β β β ββ build_hf_gpt_transformer() (20 lines) β ββ Builds HuggingFace GPT2 with custom embeddings β ββ External Dependencies: torch, transformers, indextts.gpt modules ββ Critical Inference Parameters: ββ Temperature control for generation ββ Top-k/top-p sampling ββ Classifier-free guidance scale ββ Generation length limits /home/user/IndexTTS-Rust/indextts/gpt/conformer_encoder.py (520 LINES) ββ ββ Purpose: Conformer-based speaker conditioning encoder ββ Key Classes: β ββ ConformerEncoder (main) β β ββ Modules: β β β ββ Subsampling layer (Conv2d) β β β ββ Positional encoding β β β ββ Conformer blocks β β β ββ Layer normalization β β β ββ Optional projection layer β β β β β ββ Configuration Parameters: β β β ββ input_size: 1024 (mel spectrogram bins) β β β ββ output_size: depends on config β β β ββ linear_units: hidden dim for FFN β β β ββ attention_heads: 8 β β β ββ num_blocks: 4 β β β ββ input_layer: "linear" or "conv2d" β β β β β ββ Architecture: Conv β Pos Enc β [Conformer Block] * N β LayerNorm β β β ββ ConformerBlock (80+ lines) β β ββ Residual connections β β ββ FFN β Attention β Conv β FFN structure β β ββ Feed-forward network (2-layer with dropout) β β ββ Multi-head self-attention β β ββ Convolution module (depthwise) β β ββ Layer normalization β β β ββ ConvolutionModule (50 lines) β β ββ Pointwise Conv 1x1 β β ββ Depthwise Conv with kernel_size (e.g., 15) β β ββ Batch normalization or layer normalization β β ββ Activation (ReLU/SiLU) β β ββ Projection β β β ββ PositionwiseFeedForward (15 lines) β β ββ Dense layer (idim β hidden) β β ββ Activation (ReLU) β β ββ Dropout β β ββ Dense layer (hidden β idim) β β β ββ MultiHeadedAttention (custom) β ββ Scaled dot-product attention β ββ Multiple heads β ββ Optional relative position bias β ββ External Dependencies: torch, custom conformer modules ββ Use Case: Processing mel spectrogram to extract speaker features /home/user/IndexTTS-Rust/indextts/gpt/perceiver.py (317 LINES) ββ ββ Purpose: Perceiver resampler for attention pooling ββ Key Classes: β ββ PerceiverResampler (250+ lines) β β ββ Architecture: β β β ββ Learnable latent queries β β β ββ Cross-attention layers β β β ββ Feed-forward networks β β β ββ Layer normalization β β β β β ββ Parameters: β β β ββ dim: 512 (embedding dimension) β β β ββ dim_context: 512 (context dimension) β β β ββ num_latents: 32 (number of latent queries) β β β ββ num_latent_channels: 64 β β β ββ num_layers: 6 β β β ββ ff_mult: 4 (FFN expansion) β β β ββ heads: 8 β β β β β ββ Key Methods: β β β ββ forward() - Attend and pool β β β ββ _cross_attend_block() - Single cross-attention layer β β β β β ββ Cross-Attention Mechanism: β β ββ Queries: Learnable latents β β ββ Keys/Values: Input context β β ββ Output: Pooled features (num_latents Γ dim) β β ββ FFN projection for dimension mixing β β β ββ FeedForward (15 lines) β ββ Dense (dim β hidden) β ββ GELU activation β ββ Dense (hidden β dim) β ββ External Dependencies: torch, einsum operations ββ Use Case: Pool conditioning encoder output to fixed-size representation VOCODER & AUDIO SYNTHESIS FILES βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ /home/user/IndexTTS-Rust/indextts/BigVGAN/models.py (1000+ LINES) βββ ββ Purpose: BigVGAN neural vocoder for mel-to-audio conversion ββ Key Classes: β ββ BigVGAN (400+ lines) β β ββ Architecture: β β β ββ Initial Conv1d (80 mel bins β 192 channels) β β β ββ Upsampling layers (transposed conv) β β β ββ AMP blocks (anti-aliased multi-period) β β β ββ Final Conv1d (channels β 1 waveform) β β β ββ Tanh activation for output β β β β β ββ Upsampling: 4x β 8x β 8x β 4x (256x total) β β β ββ Maps from 22050 Hz mel frames to audio samples β β β ββ Kernel sizes: [16, 16, 4, 4] β β β ββ Padding: [6, 6, 2, 2] β β β β β ββ Parameters: β β β ββ num_mels: 80 β β β ββ num_freq: 513 β β β ββ num_mels: 80 β β β ββ n_fft: 1024 β β β ββ hop_size: 256 β β β ββ win_size: 1024 β β β ββ sampling_rate: 22050 β β β ββ freq_min: 0 β β β ββ freq_max: None β β β ββ use_cuda_kernel: bool β β β β β ββ Key Methods: β β β ββ forward() - Mel β audio waveform β β β ββ from_pretrained() - Load from HuggingFace β β β ββ remove_weight_norm() - Remove spectral normalization β β β ββ eval() - Set to evaluation mode β β β β β ββ Special Features: β β ββ Weight normalization for training stability β β ββ Spectral normalization option β β ββ CUDA kernel support for activation functions β β ββ Snake/SnakeBeta activation (periodic) β β ββ Anti-aliasing filters for high-quality upsampling β β β ββ AMPBlock1 (50 lines) β β ββ Architecture: Conv1d Γ 2 with activations β β ββ Multiple dilation patterns [1, 3, 5] β β ββ Residual connections β β ββ Activation1d wrapper for anti-aliasing β β ββ Weight normalization β β β ββ AMPBlock2 (40 lines) β β ββ Similar to AMPBlock1 but simpler β β ββ Dilation patterns [1, 3] β β ββ Residual connections β β β ββ Activation1d (custom, from alias_free_activation/) β β ββ Applies activation function (Snake/SnakeBeta) β β ββ Optional anti-aliasing filter β β ββ Optional CUDA kernel for efficiency β β β ββ Snake Activation (from activations.py) β β ββ Formula: x + (1/alpha) * sinΒ²(alpha * x) β β ββ Periodic nonlinearity β β ββ Learnable alpha parameter β β β ββ SnakeBeta Activation (from activations.py) β ββ More complex periodic activation β ββ Improved harmonic modeling β ββ External Dependencies: torch, scipy, librosa ββ Model Size: ~100 MB (pretrained weights) /home/user/IndexTTS-Rust/indextts/s2mel/modules/audio.py (83 LINES) ββ Purpose: Mel-spectrogram computation (DSP) ββ Key Functions: β ββ load_wav() - Load WAV file with scipy β ββ mel_spectrogram() - Compute mel spectrogram β β ββ Parameters: β β β ββ y: waveform tensor β β β ββ n_fft: 1024 β β β ββ num_mels: 80 β β β ββ sampling_rate: 22050 β β β ββ hop_size: 256 β β β ββ win_size: 1024 β β β ββ fmin: 0 β β β ββ fmax: None or 8000 β β β β β ββ Process: β β β 1. Pad input with reflect padding β β β 2. Compute STFT (Short-Time Fourier Transform) β β β 3. Convert to magnitude spectrogram β β β 4. Apply mel filterbank (librosa) β β β 5. Apply dynamic range compression (log) β β β ββ Output: [1, 80, T] tensor β β β β β ββ Caching: β β ββ Caches mel filterbank matrices β β ββ Caches Hann windows β β ββ Device-specific caching β β β ββ dynamic_range_compression() - Log compression β ββ dynamic_range_decompression() - Inverse β ββ spectral_normalize/denormalize() β ββ Critical DSP Parameters: β ββ STFT Window: Hann window β ββ FFT Size: 1024 β ββ Hop Size: 256 (11.6 ms at 22050 Hz) β ββ Mel Bins: 80 (perceptual scale) β ββ Min Freq: 0 Hz β ββ Max Freq: Variable (8000 Hz or Nyquist) β ββ External Dependencies: torch, librosa, scipy SEMANTIC CODEC & FEATURE EXTRACTION FILES βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ /home/user/IndexTTS-Rust/indextts/utils/maskgct_utils.py (250 LINES) ββ Purpose: Build and manage semantic codecs ββ Key Functions: β ββ build_semantic_model() β β ββ Loads: facebook/w2v-bert-2.0 model β β ββ Extracts: wav2vec 2.0 BERT embeddings β β ββ Returns: model, mean, std (for normalization) β β ββ Output: 1024-dimensional embeddings β β β ββ build_semantic_codec() β β ββ Creates: RepCodec (residual vector quantization) β β ββ Quantizes: Semantic embeddings β β ββ Returns: Codec model β β ββ Output: Discrete tokens β β β ββ build_s2a_model() β β ββ Builds: MaskGCT_S2A (semantic-to-acoustic) β β ββ Maps: Semantic codes β acoustic codes β β β ββ build_acoustic_codec() β β ββ Encoder: Encodes acoustic features β β ββ Decoder: Decodes codes β audio β β ββ Multiple codec variants β β β ββ Inference_Pipeline (class) β ββ Combines all codecs β ββ Methods: β β ββ get_emb() - Get semantic embeddings β β ββ get_scode() - Quantize to semantic codes β β ββ semantic2acoustic() - Convert codes β β ββ s2a_inference() - Full pipeline β ββ Diffusion-based generation options β ββ External Dependencies: torch, transformers, huggingface_hub ββ Pre-trained Models: ββ W2V-BERT-2.0: 614M parameters ββ MaskGCT: From amphion/MaskGCT ββ Various codec checkpoints CONFIGURATION & UTILITY FILES βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ /home/user/IndexTTS-Rust/indextts/utils/checkpoint.py (50 LINES) ββ Purpose: Load model checkpoints ββ Key Functions: β ββ load_checkpoint() - Load weights into model β ββ Device handling (CPU/GPU/XPU/MPS) ββ Supported Formats: .pth, .safetensors /home/user/IndexTTS-Rust/indextts/utils/arch_util.py ββ Purpose: Architecture utility modules ββ Key Classes: β ββ AttentionBlock - Generic attention layer ββ Used in: Conditioning encoder, other modules /home/user/IndexTTS-Rust/indextts/utils/xtransformers.py (1,600 LINES) ββ Purpose: Extended transformer utilities ββ Key Components: β ββ Advanced attention mechanisms β ββ Relative position bias β ββ Cross-attention patterns β ββ Various position encoding schemes ββ Used in: GPT model, encoders TESTING FILES βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ /home/user/IndexTTS-Rust/tests/regression_test.py ββ Test Cases: β ββ Chinese text with pinyin tones (ζ XUAN4) β ββ English text β ββ Mixed Chinese-English β ββ Long-form text with multiple sentences β ββ Named entities (Joseph Gordon-Levitt) β ββ Chinese names (ηΊ¦η倫·ι«η»-θ±η»΄ηΉ) β ββ Extended passages for robustness ββ Inference Modes: β ββ Single inference (infer) β ββ Fast inference (infer_fast) ββ Output: WAV files in outputs/ directory /home/user/IndexTTS-Rust/tests/padding_test.py ββ Test Scenarios: β ββ Variable length inputs β ββ Batch processing β ββ Edge cases β ββ Padding handling ββ Purpose: Ensure robust padding mechanics βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ KEY ALGORITHMS SUMMARY: 1. TEXT PROCESSING: - Regex-based pattern matching for pinyin/names - Character-level CJK tokenization - SentencePiece BPE encoding - Language detection (Chinese vs English) 2. FEATURE EXTRACTION: - W2V-BERT semantic embeddings (1024-dim) - RepCodec quantization - Mel-spectrogram (STFT-based, 80-dim) - CAMPPlus speaker embeddings (192-dim) 3. SEQUENCE GENERATION: - GPT-based autoregressive generation - Conformer speaker conditioning - Perceiver pooling for attention - Classifier-free guidance (optional) - Temperature/top-k/top-p sampling 4. AUDIO SYNTHESIS: - Transposed convolution upsampling (256x) - Anti-aliased activation functions - Residual connections - Weight/spectral normalization 5. EMOTION CONTROL: - 8-dimensional emotion vectors - Text-based emotion detection (via Qwen) - Audio-based emotion extraction - Emotion matrix interpolation βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ |