File size: 40,378 Bytes
f6c1e7e |
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 |
# coding=utf-8
# Copyright 2025 SparkAudio & The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Processor class for SparkTTS. Combines text tokenization and audio feature extraction/processing.
"""
import os # Needed for save_pretrained
import re # For decoding
import torch
import numpy as np
import soundfile as sf # For audio loading
import soxr # For resampling
from pathlib import Path
from typing import Optional, Union, List, Dict, Tuple, Any
from transformers.processing_utils import ProcessorMixin
from transformers.tokenization_utils_base import BatchEncoding # Return type hint
from transformers.feature_extraction_utils import BatchFeature # Return type hint
from transformers.models.auto.tokenization_auto import AutoTokenizer
from transformers.models.wav2vec2.feature_extraction_wav2vec2 import Wav2Vec2FeatureExtractor
from transformers.utils import logging, PushToHubMixin # Added PushToHubMixin
from numpy.lib.stride_tricks import sliding_window_view
import soxr
import soundfile
import random
# Import custom config if needed for defaults
from .configuration_spark_tts import SparkTTSConfig
logger = logging.get_logger(__name__)
# =============================================================================
# >> START: PASTE CODE FROM sparktts/utils/* HERE <<
# =============================================================================
# IMPORTANT: Utility functions needed for processing (audio loading, token parsing)
# must be defined or imported here.
# --- Paste sparktts/utils/audio.py content here ---
def audio_volume_normalize(audio: np.ndarray, coeff: float = 0.2) -> np.ndarray:
"""
Normalize the volume of an audio signal.
Parameters:
audio (numpy array): Input audio signal array.
coeff (float): Target coefficient for normalization, default is 0.2.
Returns:
numpy array: The volume-normalized audio signal.
"""
# Sort the absolute values of the audio signal
temp = np.sort(np.abs(audio))
# If the maximum value is less than 0.1, scale the array to have a maximum of 0.1
if temp[-1] < 0.1:
scaling_factor = max(
temp[-1], 1e-3
) # Prevent division by zero with a small constant
audio = audio / scaling_factor * 0.1
# Filter out values less than 0.01 from temp
temp = temp[temp > 0.01]
L = temp.shape[0] # Length of the filtered array
# If there are fewer than or equal to 10 significant values, return the audio without further processing
if L <= 10:
return audio
# Compute the average of the top 10% to 1% of values in temp
volume = np.mean(temp[int(0.9 * L) : int(0.99 * L)])
# Normalize the audio to the target coefficient level, clamping the scale factor between 0.1 and 10
audio = audio * np.clip(coeff / volume, a_min=0.1, a_max=10)
# Ensure the maximum absolute value in the audio does not exceed 1
max_value = np.max(np.abs(audio))
if max_value > 1:
audio = audio / max_value
return audio
def load_audio(
adfile: Path,
sampling_rate: int = None,
length: int = None,
volume_normalize: bool = False,
segment_duration: int = None,
) -> np.ndarray:
r"""Load audio file with target sampling rate and lsength
Args:
adfile (Path): path to audio file.
sampling_rate (int, optional): target sampling rate. Defaults to None.
length (int, optional): target audio length. Defaults to None.
volume_normalize (bool, optional): whether perform volume normalization. Defaults to False.
segment_duration (int): random select a segment with duration of {segment_duration}s.
Defualt to None which means the whole audio will be used.
Returns:
audio (np.ndarray): audio
"""
audio, sr = soundfile.read(adfile)
if len(audio.shape) > 1:
audio = audio[:, 0]
if sampling_rate is not None and sr != sampling_rate:
audio = soxr.resample(audio, sr, sampling_rate, quality="VHQ")
sr = sampling_rate
if segment_duration is not None:
seg_length = int(sr * segment_duration)
audio = random_select_audio_segment(audio, seg_length)
# Audio volume normalize
if volume_normalize:
audio = audio_volume_normalize(audio)
# check the audio length
if length is not None:
assert abs(audio.shape[0] - length) < 1000
if audio.shape[0] > length:
audio = audio[:length]
else:
audio = np.pad(audio, (0, int(length - audio.shape[0])))
return audio
def random_select_audio_segment(audio: np.ndarray, length: int) -> np.ndarray:
"""get an audio segment given the length
Args:
audio (np.ndarray):
length (int): audio length = sampling_rate * duration
"""
if audio.shape[0] < length:
audio = np.pad(audio, (0, int(length - audio.shape[0])))
start_index = random.randint(0, audio.shape[0] - length)
end_index = int(start_index + length)
return audio[start_index:end_index]
def get_ref_clip(wav: np.ndarray, config) -> np.ndarray: # Needs access to config attributes
"""Get reference audio clip for speaker embedding."""
# Make sure config has sample_rate, ref_segment_duration, latent_hop_length
if not all(hasattr(config, attr) for attr in ['sample_rate', 'ref_segment_duration', 'latent_hop_length']):
raise AttributeError("Config object missing required attributes for get_ref_clip")
ref_segment_length = (
int(config.sample_rate * config.ref_segment_duration)
// config.latent_hop_length
* config.latent_hop_length
)
wav_length = len(wav)
if ref_segment_length > wav_length:
wav = np.tile(wav, ref_segment_length // wav_length + 1)
return wav[:ref_segment_length]
# --- Paste sparktts/utils/token_parser.py content here ---
TASK_TOKEN_MAP = {
"vc": "<|task_vc|>",
"tts": "<|task_tts|>",
"asr": "<|task_asr|>",
"s2s": "<|task_s2s|>",
"t2s": "<|task_t2s|>",
"understand": "<|task_understand|>",
"caption": "<|task_cap|>",
"controllable_tts": "<|task_controllable_tts|>",
"prompt_tts": "<|task_prompt_tts|>",
"speech_edit": "<|task_edit|>",
}
LEVELS_MAP = {
"very_low": 0,
"low": 1,
"moderate": 2,
"high": 3,
"very_high": 4,
}
LEVELS_MAP_UI = {
1: 'very_low',
2: 'low',
3: 'moderate',
4: 'high',
5: 'very_high'
}
GENDER_MAP = {
"female": 0,
"male": 1,
}
AGE_MAP = {"Child": 0, "Teenager": 1, "Youth-Adult": 2, "Middle-aged": 3, "Elderly": 4}
EMO_MAP = {
"UNKNOWN": 0,
"NEUTRAL": 1,
"ANGRY": 2,
"HAPPY": 3,
"SAD": 4,
"FEARFUL": 5,
"DISGUSTED": 6,
"SURPRISED": 7,
"SARCASTIC": 8,
"EXCITED": 9,
"SLEEPY": 10,
"CONFUSED": 11,
"EMPHASIS": 12,
"LAUGHING": 13,
"SINGING": 14,
"WORRIED": 15,
"WHISPER": 16,
"ANXIOUS": 17,
"NO-AGREEMENT": 18,
"APOLOGETIC": 19,
"CONCERNED": 20,
"ENUNCIATED": 21,
"ASSERTIVE": 22,
"ENCOURAGING": 23,
"CONTEMPT": 24,
}
class TokenParser:
"""Turn label to special token"""
def __init__(self):
pass
"""Parse the attributes of a person."""
def __init__(self):
pass
@staticmethod
def age(age: str) -> str:
"""Turn age token."""
age_id = AGE_MAP[age]
return f"<|age_{age_id}|>"
@staticmethod
def gender(gender: str) -> str:
"""Turn gender token."""
gender_id = GENDER_MAP[gender]
return f"<|gender_{gender_id}|>"
@staticmethod
def mel_value(mel: int):
"""Turn special token of mel scale pitch."""
mel = max(0, int(mel))
mel = min(1000, int(mel))
return f"<|pitch_value_{mel}|>"
@staticmethod
def mel_level(level: str):
"""Turn special token of mel level."""
level_tag = LEVELS_MAP[level]
return f"<|pitch_label_{level_tag}|>"
@staticmethod
def pitch_var_value(pitch_std: int):
"""Turn special token of pitch_std value."""
assert isinstance(pitch_std, int)
pitch_std = max(0, int(pitch_std))
pitch_std = min(10, int(pitch_std))
return f"<|pitch_var_value_{pitch_std}|>"
@staticmethod
def pitch_var_level(level: str):
"""Turn special token of pitch std level."""
level_tag = LEVELS_MAP[level]
return f"<|pitch_var_label_{level_tag}|>"
@staticmethod
def loudness_value(loudness: int):
"""Turn special toak of loudness value [0, 30]"""
assert loudness >= 0
loudness = max(0, int(loudness))
loudness = min(30, int(loudness))
return f"<|loudness_value_{loudness}|>"
@staticmethod
def loudness_level(level: str):
"""Turn special token of loudness level."""
level_tag = LEVELS_MAP[level]
return f"<|loudness_label_{level_tag}|>"
@staticmethod
def speed_value(speed: int):
"""Turn special token of speed value."""
speed = max(0, int(speed))
speed = min(10, int(speed))
return f"<|speed_value_{speed}|>"
@staticmethod
def speed_level(level: str):
"""Turn special token of speed level."""
level_tag = LEVELS_MAP[level]
return f"<|speed_label_{level_tag}|>"
@staticmethod
def task(task: str) -> str:
"""Turn special token of task."""
assert task in TASK_TOKEN_MAP.keys()
return TASK_TOKEN_MAP[task]
@staticmethod
def emotion(emotion: str):
emo_id = EMO_MAP[emotion]
return f"<|emotion_{emo_id}|>"
# =============================================================================
# >> END: PASTE CODE FROM sparktts/utils/* HERE <<
# =============================================================================
class SparkTTSProcessor(ProcessorMixin, PushToHubMixin): # Added PushToHubMixin
r"""
Constructs a SparkTTS processor which wraps a text tokenizer and relevant audio processing logic.
Args:
tokenizer ([`PreTrainedTokenizer`]):
An instance of [`PreTrainedTokenizer`]. This handles the text tokenization for the LLM.
feature_extractor ([`Wav2Vec2FeatureExtractor`]):
An instance of [`Wav2Vec2FeatureExtractor`]. Although Wav2Vec2 features are extracted
within the model's `tokenize_audio`, the extractor's configuration (like sampling rate)
is useful, and it aligns with the ProcessorMixin pattern.
config ([`SparkTTSConfig`], *optional*):
An instance of [`SparkTTSConfig`] to access configuration parameters like sample rate.
"""
attributes = ["tokenizer", "feature_extractor"]
tokenizer_class = "AutoTokenizer"
feature_extractor_class = "Wav2Vec2FeatureExtractor" # Keep for consistency
def __init__(self, tokenizer, feature_extractor, config: Optional[SparkTTSConfig] = None, **kwargs):
super().__init__(tokenizer=tokenizer, feature_extractor=feature_extractor, **kwargs)
self.model = None
self.config = config
# Set sampling rate
if config and hasattr(config, 'sample_rate'):
self.sampling_rate = config.sample_rate
elif feature_extractor and hasattr(feature_extractor, 'sampling_rate'):
self.sampling_rate = feature_extractor.sampling_rate
else:
self.sampling_rate = 16000
logger.warning(f"Could not determine sampling rate. Defaulting to {self.sampling_rate} Hz.")
# # Ensure tokenizer pad token
# if self.tokenizer.pad_token is None:
# if self.tokenizer.eos_token is not None:
# logger.warning("Tokenizer does not have a pad token. Setting pad_token to eos_token.")
# self.tokenizer.pad_token = self.tokenizer.eos_token
# else:
# logger.warning("Tokenizer lacks pad and eos token. Adding default pad token '<|pad|>'.")
# self.tokenizer.add_special_tokens({'pad_token': '<|pad|>'})
def link_model(self, model):
"""Links the processor to a SparkTTSModel instance for audio processing calls."""
if not hasattr(model, 'tokenize_audio') or not hasattr(model, 'detokenize_audio'):
raise TypeError("The provided model instance does not have the required 'tokenize_audio' and 'detokenize_audio' methods.")
if not hasattr(model, 'config'):
logger.warning("Linked model does not have a 'config' attribute. Some processor functionalities might rely on it.")
self.model = model
logger.info("SparkTTSModel successfully linked to the processor.")
# Update sampling rate based on linked model's config if available
if hasattr(model, 'config') and hasattr(model.config, 'sample_rate'):
if self.sampling_rate != model.config.sample_rate:
logger.info(f"Updating processor sampling rate from {self.sampling_rate} to {model.config.sample_rate} based on linked model config.")
self.sampling_rate = model.config.sample_rate
# Also update feature extractor sampling rate if it differs
if hasattr(self, 'feature_extractor') and self.feature_extractor.sampling_rate != model.config.sample_rate:
logger.info(f"Updating feature_extractor sampling rate from {self.feature_extractor.sampling_rate} to {model.config.sample_rate}.")
self.feature_extractor.sampling_rate = model.config.sample_rate
def __call__(
self,
text: str,
prompt_speech_path: Optional[Union[str, Path]] = None,
prompt_text: Optional[str] = None,
gender: Optional[str] = None,
pitch: Optional[str] = None,
speed: Optional[str] = None,
return_tensors: Optional[str] = "pt",
**kwargs, # Allow passing other args like padding, truncation to tokenizer
) -> BatchEncoding:
"""
Processes the input text and optional prompt audio/control parameters into a format suitable for [`SparkTTSModel`].
Args:
text (`str`):
The main text to be synthesized.
prompt_speech_path (`str` or `Path`, *optional*):
Path to the prompt audio file for voice cloning. Required if `gender` is not set.
prompt_text (`str`, *optional*):
Transcript of the prompt audio. Used only in voice cloning mode.
gender (`str`, *optional*):
Target gender ("male" or "female") for controllable synthesis. If set, enables control mode.
pitch (`str`, *optional*):
Target pitch level ("very_low", "low", "moderate", "high", "very_high") for control mode. Required if `gender` is set.
speed (`str`, *optional*):
Target speed level ("very_low", "low", "moderate", "high", "very_high") for control mode. Required if `gender` is set.
return_tensors (`str`, *optional*, defaults to `"pt"`):
If set, will return tensors instead of list of python integers. Only "pt" (PyTorch) is supported currently.
**kwargs:
Additional arguments passed to the underlying tokenizer's `__call__` method.
Returns:
[`BatchEncoding`]: A dictionary containing the `input_ids` and `attention_mask` for the LLM.
In voice cloning mode, it also includes `global_token_ids_prompt` (torch.Tensor) representing the
global tokens extracted from the prompt audio.
"""
global_token_ids_prompt = None # Initialize
# Determine mode: Control TTS or Voice Cloning (Prompt TTS)
is_control_mode = gender is not None
is_cloning_mode = prompt_speech_path is not None and not is_control_mode
if is_control_mode:
# --- Controllable TTS Prompt Construction ---
if not all([pitch, speed]):
raise ValueError("For controllable TTS, 'gender', 'pitch', and 'speed' must all be provided.")
if prompt_speech_path is not None:
logger.warning("`prompt_speech_path` provided but ignored because `gender` is set (controllable TTS mode).")
if not all(k in GENDER_MAP for k in [gender]): # Basic check
raise ValueError(f"Invalid gender provided: {gender}. Must be one of {list(GENDER_MAP.keys())}")
if not all(k in LEVELS_MAP for k in [pitch, speed]): # Basic check
raise ValueError(f"Invalid pitch or speed level provided. Must be one of {list(LEVELS_MAP.keys())}")
gender_id = GENDER_MAP[gender]
pitch_level_id = LEVELS_MAP[pitch]
speed_level_id = LEVELS_MAP[speed]
pitch_label_tokens = f"<|pitch_label_{pitch_level_id}|>"
speed_label_tokens = f"<|speed_label_{speed_level_id}|>"
gender_tokens = f"<|gender_{gender_id}|>"
attribute_tokens = "".join([gender_tokens, pitch_label_tokens, speed_label_tokens])
prompt_list = [
TASK_TOKEN_MAP["controllable_tts"],
"<|start_content|>",
text,
"<|end_content|>",
"<|start_style_label|>",
attribute_tokens,
"<|end_style_label|>",
]
prompt_string = "".join(prompt_list)
elif is_cloning_mode:
# --- Voice Cloning Prompt Construction ---
if self.model is None:
raise RuntimeError("Processor must be linked to a SparkTTSModel instance via `processor.link_model(model)` before performing voice cloning.")
prompt_speech_path = Path(prompt_speech_path) # Ensure it's a Path object
if not prompt_speech_path.exists():
raise FileNotFoundError(f"Prompt audio file not found: {prompt_speech_path}")
# Load and process prompt audio
try:
model_config = self.model.config if self.model and hasattr(self.model, 'config') else self.config
if model_config is None:
raise ValueError("Configuration not available in processor or linked model.")
# Load main wav
wav = load_audio(
prompt_speech_path,
sampling_rate=self.sampling_rate,
volume_normalize=getattr(model_config, 'volume_normalize', True), # Use getattr for safety
)
# Get reference clip
wav_ref_np = get_ref_clip(wav, model_config) # Pass config object
wav_ref = torch.from_numpy(wav_ref_np).unsqueeze(0).float()
wav_tensor = torch.from_numpy(wav).unsqueeze(0).float()
# Tokenize using the linked model's method
# Assuming tokenize_audio returns tensors with batch dim 1: [1, N_global], [1, N_semantic]
global_tokens_tensor, semantic_tokens_tensor = self.model.tokenize_audio(wav_tensor, wav_ref)
# Store the global tokens tensor (with batch dim) for the output dict
global_token_ids_prompt = global_tokens_tensor # Keep batch dim [1, N_global]
# Convert tensors to lists of ints for string formatting
global_token_list = global_tokens_tensor.squeeze().tolist() # Remove batch dim -> list
semantic_token_list = semantic_tokens_tensor.squeeze().tolist() # Remove batch dim -> list
except Exception as e:
logger.error(f"Error processing prompt audio {prompt_speech_path}: {e}")
import traceback
traceback.print_exc()
raise
# ==============================================================
# CORRECTED TOKEN STRING FORMATTING
# ==============================================================
# Create individual token strings for each ID
global_tokens_str = "".join([f"<|bicodec_global_{gid}|>" for gid in global_token_list])
semantic_tokens_str = "".join([f"<|bicodec_semantic_{sid}|>" for sid in semantic_token_list])
# ==============================================================
# Construct prompt list based on presence of prompt_text
if prompt_text is not None and prompt_text.strip(): # Check if prompt_text is meaningful
logger.info("Using prompt text in voice cloning prompt.")
prompt_list = [
TASK_TOKEN_MAP["tts"], # Or maybe TASK_TOKEN_MAP["prompt_tts"]? Check original logic. Assuming "tts".
"<|start_content|>",
prompt_text, # Transcript first
text, # Then target text
"<|end_content|>",
"<|start_global_token|>",
global_tokens_str,
"<|end_global_token|>",
"<|start_semantic_token|>",
semantic_tokens_str,
# "<|end_semantic_token|>", # Original code didn't have this marker here
]
else:
# Simpler prompt without semantic tokens if no transcript provided
logger.info("No prompt text provided, using text-only voice cloning prompt.")
prompt_list = [
TASK_TOKEN_MAP["tts"], # Or maybe TASK_TOKEN_MAP["prompt_tts"]?
"<|start_content|>",
text, # Only target text
"<|end_content|>",
"<|start_global_token|>",
global_tokens_str,
"<|end_global_token|>",
]
prompt_string = "".join(prompt_list)
logger.debug(f"Generated prompt string (cloning): {prompt_string[:200]}...") # Log start of prompt
else:
raise ValueError("Invalid input combination. Either provide `prompt_speech_path` for cloning or (`gender`, `pitch`, `speed`) for control.")
# --- Tokenize the final prompt string ---
# print(f"Tokenizing prompt: {prompt_string}")
inputs = self.tokenizer(
prompt_string,
return_tensors=return_tensors,
padding=kwargs.get("padding", False), # Often False for generation prompts unless batching > 1
truncation=kwargs.get("truncation", True),
max_length=kwargs.get("max_length", self.tokenizer.model_max_length),
add_special_tokens=kwargs.get("add_special_tokens", True), # Usually True unless handled manually
return_attention_mask=kwargs.get("return_attention_mask", True), # Need attention mask
**{k: v for k, v in kwargs.items() if k not in ["padding", "truncation", "max_length", "add_special_tokens", "return_attention_mask"]}
)
logger.debug(f"Tokenized input_ids shape: {inputs['input_ids'].shape}")
# Add the prompt's global tokens (as tensor with batch dim) to the output if in cloning mode
if is_cloning_mode and global_token_ids_prompt is not None:
if return_tensors == "pt":
inputs["global_token_ids_prompt"] = global_token_ids_prompt # Already has batch dim [1, N_global]
else:
# Handle non-tensor return if necessary
inputs["global_token_ids_prompt"] = global_token_ids_prompt.tolist()
return inputs
def decode(
self,
generated_ids: torch.Tensor,
global_token_ids_prompt: Optional[torch.Tensor] = None,
input_ids_len: Optional[int] = None,
skip_special_tokens: bool = True,
) -> Dict[str, Any]:
"""
Decodes the generated token IDs from [`SparkTTSModel`] into an audio waveform.
Args:
generated_ids (`torch.Tensor`):
Tensor of token IDs generated by `model.generate()`, including the input prompt part. Shape [B, seq_len].
global_token_ids_prompt (`torch.Tensor`, *optional*):
The global tokens extracted from the prompt audio during the `__call__` step (for voice cloning).
Shape [B, N_global]. Required if the generation was for voice cloning.
input_ids_len (`int`, *optional*):
The length of the original input prompt `input_ids` fed to `model.generate()`. Required to
correctly isolate the newly generated tokens.
skip_special_tokens (`bool`, *optional*, defaults to `True`):
Whether to skip special tokens during the text decoding step (used to extract audio tokens).
Returns:
Dict[str, Any]: A dictionary containing:
- "audio": The decoded audio waveform as a NumPy array. Shape [T_audio] (if B=1) or [B, T_audio].
- "sampling_rate": The sampling rate of the audio.
"""
if self.model is None:
raise RuntimeError("Processor must be linked to a SparkTTSModel instance via `processor.link_model(model)` before decoding.")
if input_ids_len is None:
raise ValueError("`input_ids_len` (length of the prompt input_ids) must be provided for decoding.")
# --- Isolate generated part and decode text ---
# Assumes generated_ids has shape [B, full_seq_len]
# Handle case where generated sequence is shorter than prompt (shouldn't happen with max_new_tokens > 0)
if generated_ids.shape[1] < input_ids_len:
logger.warning(f"Generated sequence length ({generated_ids.shape[1]}) is shorter than input prompt length ({input_ids_len}). Decoding might be incorrect.")
output_only_ids = generated_ids[:, input_ids_len:] # Will be empty if equal
else:
output_only_ids = generated_ids[:, input_ids_len:]
# Decode the generated part to find audio tokens
# Need to handle batch decoding if B > 1
# print("decode token", self.tokenizer.batch_decode(output_only_ids, skip_special_tokens=False))
decoded_texts = self.tokenizer.batch_decode(output_only_ids, skip_special_tokens=skip_special_tokens)
# --- Extract Audio Tokens ---
# Handle batch processing correctly
batch_size = generated_ids.shape[0]
all_semantic_ids = []
all_global_tokens = []
successful_indices = [] # Keep track of which batch items were successful
for i in range(batch_size):
decoded_text = decoded_texts[i]
current_semantic_ids = None
current_global_tokens = None
# Extract semantic tokens
try:
pred_semantic_indices = [int(token) for token in re.findall(r"bicodec_semantic_(\d+)", decoded_text)]
if not pred_semantic_indices:
logger.warning(f"Batch item {i}: No semantic tokens found in decoded text: '{decoded_text[:200]}...'")
continue # Skip this item
current_semantic_ids = torch.tensor(pred_semantic_indices).long() # Shape [N_semantic]
except Exception as e:
logger.error(f"Batch item {i}: Error parsing semantic tokens from: '{decoded_text[:200]}...'. Error: {e}")
continue # Skip this item
# Determine global tokens
if global_token_ids_prompt is not None:
# Cloning mode: Use the provided prompt global tokens for this batch item
if global_token_ids_prompt.shape[0] != batch_size:
raise ValueError(f"Batch size mismatch: generated_ids has {batch_size}, but global_token_ids_prompt has {global_token_ids_prompt.shape[0]}.")
current_global_tokens = global_token_ids_prompt[i] # Shape [N_global]
else:
# Control mode: Extract global tokens from the generated text
try:
pred_global_indices = [int(token) for token in re.findall(r"bicodec_global_(\d+)", decoded_text)]
if not pred_global_indices:
logger.warning(f"Batch item {i}: No global tokens found in decoded text for control mode: '{decoded_text[:200]}...'")
continue # Skip this item
current_global_tokens = torch.tensor(pred_global_indices).long() # Shape [N_global]
except Exception as e:
logger.error(f"Batch item {i}: Error parsing global tokens from: '{decoded_text[:200]}...'. Error: {e}")
continue # Skip this item
# If both tokens extracted successfully
all_semantic_ids.append(current_semantic_ids)
all_global_tokens.append(current_global_tokens)
successful_indices.append(i)
if not successful_indices:
logger.error("Failed to extract audio tokens for any item in the batch.")
return {"audio": np.array([], dtype=np.float32), "sampling_rate": self.sampling_rate}
# Pad sequences to the max length within the successful batch items for batch detokenization
# Note: BiCodec might not support batching if sequences have different lengths. Check its implementation.
# Assuming BiCodec *can* handle batches if padded (or if lengths are naturally equal).
# This padding might be unnecessary if BiCodec handles variable lengths or if B=1 anyway.
# For now, let's assume B=1 was handled correctly and skip complex padding.
if batch_size > 1 and len(successful_indices) < batch_size:
logger.warning(f"Only successfully decoded {len(successful_indices)} out of {batch_size} batch items.")
# Further processing might need to handle only the successful items.
# Let's proceed assuming B=1 or BiCodec handles batches appropriately.
# Stack the successful tokens.
try:
# Need to ensure tensors have the same length before stacking if BiCodec requires it.
# If BiCodec handles variable length, stacking might not be needed, just loop and call detokenize.
# Let's assume B=1 for simplicity of the example, matching original code's likely behavior.
if len(successful_indices) != 1:
raise NotImplementedError("Batch decoding (B > 1) requires verification of BiCodec's batch handling and potentially padding.")
final_semantic_ids = all_semantic_ids[0].unsqueeze(0) # Add batch dim [1, N_semantic]
final_global_tokens = all_global_tokens[0].unsqueeze(0) # Add batch dim [1, N_global]
except IndexError: # Should not happen if successful_indices is not empty
logger.error("Internal error during token batch preparation.")
return {"audio": np.array([], dtype=np.float32), "sampling_rate": self.sampling_rate}
# --- Detokenize Audio ---
try:
# Call the linked model's detokenize method
# print(f"DEBUG: Detokenizing audio with global tokens {final_global_tokens.shape}, semantic tokens {final_semantic_ids.shape}")
output_wav = self.model.detokenize_audio(final_global_tokens, final_semantic_ids)
# detokenize_audio now returns numpy array float32 in [-1, 1]
# Optional: Double-check dtype here if needed, but should be handled by detokenize_audio now
# if output_wav.dtype != np.float32:
# logger.warning(f"Audio dtype after detokenize is {output_wav.dtype}. Converting to float32.")
# output_wav = output_wav.astype(np.float32)
# output_wav = np.clip(output_wav, -1.0, 1.0) # Clipping done in detokenize_audio
except Exception as e:
logger.error(f"Error during audio detokenization: {e}")
import traceback
traceback.print_exc()
raise RuntimeError("Audio detokenization failed.") from e
return {"audio": output_wav, "sampling_rate": self.sampling_rate}
@classmethod
def from_pretrained(
cls,
pretrained_model_name_or_path: Union[str, os.PathLike],
cache_dir: Optional[Union[str, os.PathLike]] = None,
force_download: bool = False,
local_files_only: bool = False,
token: Optional[Union[str, bool]] = None,
revision: str = "main",
trust_remote_code: bool = False, # Allow passing this, needed for config potentially
**kwargs,
):
r"""
Instantiate a SparkTTSProcessor from pretrained components.
"""
# Pop specific kwargs for this method
config = kwargs.pop("config", None) # Allow passing config explicitly
# --- 1. Load Config (to find component paths) ---
# We need the config even if the processor doesn't store it permanently,
# just to find where the tokenizer/feature_extractor live.
loaded_config = None
if not isinstance(config, SparkTTSConfig):
try:
# Load the specific config class
loaded_config = SparkTTSConfig.from_pretrained(
pretrained_model_name_or_path,
cache_dir=cache_dir,
force_download=force_download,
local_files_only=local_files_only,
token=token,
revision=revision,
trust_remote_code=trust_remote_code, # Config might be custom
**kwargs, # Pass relevant kwargs
)
except Exception as e:
logger.warning(
f"Could not load SparkTTSConfig from {pretrained_model_name_or_path}. "
f"Attempting to load components from default relative paths ('LLM', 'wav2vec2-large-xlsr-53'). Error: {e}"
)
loaded_config = None # Fallback
else:
# Config object was passed directly
loaded_config = config
# --- 2. Determine Component Paths ---
llm_tokenizer_path_or_id = "./LLM" # Default relative path
w2v_processor_path_or_id = "./wav2vec2-large-xlsr-53" # Default relative path
if loaded_config:
llm_tokenizer_path_or_id = getattr(loaded_config, 'llm_model_name_or_path', llm_tokenizer_path_or_id)
w2v_processor_path_or_id = getattr(loaded_config, 'wav2vec2_model_name_or_path', w2v_processor_path_or_id)
# The component `from_pretrained` methods handle resolving these paths/IDs
# whether they are relative subfolders of `pretrained_model_name_or_path`
# or separate Hub IDs.
# --- 3. Load Components ---
# Pass down relevant kwargs for loading components
component_loading_kwargs = {
"cache_dir": cache_dir,
"force_download": force_download,
"local_files_only": local_files_only,
"token": token,
"revision": revision,
**kwargs # Pass other user kwargs
}
try:
# Tokenizer might require trust_remote_code if its class is custom
tokenizer = AutoTokenizer.from_pretrained(
pretrained_model_name_or_path, # Main path
subfolder=llm_tokenizer_path_or_id.lstrip('./'), # Specify subfolder relative to main path
trust_remote_code=trust_remote_code,
**component_loading_kwargs
)
except Exception as e:
# Fallback: try loading directly using the path/id from config if different
if llm_tokenizer_path_or_id != "./LLM":
try:
logger.info(f"Retrying tokenizer load directly from: {llm_tokenizer_path_or_id}")
tokenizer = AutoTokenizer.from_pretrained(
llm_tokenizer_path_or_id,
trust_remote_code=trust_remote_code,
**component_loading_kwargs
)
except Exception as e2:
raise OSError(f"Could not load tokenizer using main path + subfolder or directly from '{llm_tokenizer_path_or_id}'. Error: {e2}") from e
else:
raise OSError(f"Could not load tokenizer from subfolder '{llm_tokenizer_path_or_id}' within '{pretrained_model_name_or_path}'. Error: {e}")
try:
# Feature extractor usually doesn't need trust_remote_code
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(
pretrained_model_name_or_path, # Main path
subfolder=w2v_processor_path_or_id.lstrip('./'), # Specify subfolder relative to main path
**component_loading_kwargs
)
except Exception as e:
# Fallback: try loading directly using the path/id from config if different
if w2v_processor_path_or_id != "./wav2vec2-large-xlsr-53":
try:
logger.info(f"Retrying feature extractor load directly from: {w2v_processor_path_or_id}")
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(
w2v_processor_path_or_id,
**component_loading_kwargs
)
except Exception as e2:
raise OSError(f"Could not load feature extractor using main path + subfolder or directly from '{w2v_processor_path_or_id}'. Error: {e2}") from e
else:
raise OSError(f"Could not load feature extractor from subfolder '{w2v_processor_path_or_id}' within '{pretrained_model_name_or_path}'. Error: {e}")
# --- 4. Instantiate processor ---
# Pass the potentially loaded config object (or None)
return cls(tokenizer=tokenizer, feature_extractor=feature_extractor, config=loaded_config)
def save_pretrained(
self,
save_directory: Union[str, os.PathLike],
push_to_hub: bool = False,
**kwargs,
):
"""
Save the processor's state (tokenizer and feature extractor files) to a directory.
Args:
save_directory (`str` or `os.PathLike`):
Directory where the processor files will be saved.
push_to_hub (`bool`, *optional*, defaults to `False`):
Whether or not to push your model to the Hugging Face Hub after saving it.
**kwargs:
Additional key word arguments passed along to the `push_to_hub` method.
"""
save_directory = Path(save_directory)
save_directory.mkdir(parents=True, exist_ok=True)
# Save tokenizer
self.tokenizer.save_pretrained(str(save_directory), **kwargs)
# Save feature extractor
self.feature_extractor.save_pretrained(str(save_directory), **kwargs)
# Save the main processor config (if it exists and has relevant info)
# Note: The SparkTTSConfig is usually saved with the *model*, not the processor.
# However, if the processor holds specific config needed for reloading *itself*,
# it could be saved here. Usually, relying on the model's config is sufficient.
# if self.config:
# self.config.save_pretrained(str(save_directory)) # Example if needed
logger.info(f"Processor components saved in {save_directory}")
if push_to_hub:
# Commit message and other hub kwargs can be passed via **kwargs
commit_message = kwargs.pop("commit_message", "Save processor")
return self.push_to_hub(save_directory, commit_message=commit_message, **kwargs)
return str(save_directory) # Return path consistent with Mixin |