| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | |
| 5 | import numpy as np |
| 6 | import torch |
| 7 | import torch.nn as nn |
| 8 | import torchaudio.compliance.kaldi as kaldi |
| 9 | from torch.nn.utils.rnn import pad_sequence |
| 10 | from transformers import ( |
| 11 | AutoFeatureExtractor, |
| 12 | BatchFeature, |
| 13 | ) |
| 14 | from transformers.feature_extraction_sequence_utils import SequenceFeatureExtractor |
| 15 | from transformers.processing_utils import ProcessorMixin |
| 16 | from transformers.utils import TensorType |
| 17 | |
| 18 | from vllm.logger import init_logger |
| 19 | |
| 20 | logger = init_logger(__name__) |
| 21 | |
| 22 | |
| 23 | def apply_cmvn(inputs, cmvn): # noqa |
| 24 | """Apply CMVN with mvn data.""" |
| 25 | device = inputs.device |
| 26 | # dtype = inputs.dtype |
| 27 | frame, dim = inputs.shape |
| 28 | |
| 29 | means = cmvn[0:1, :dim] |
| 30 | vars = cmvn[1:2, :dim] |
| 31 | inputs += means.to(device) |
| 32 | inputs *= vars.to(device) |
| 33 | |
| 34 | return inputs.type(torch.float32) |
| 35 | |
| 36 | |
| 37 | def apply_lfr(inputs, lfr_m, lfr_n): |
| 38 | # LFR_inputs = [] |
| 39 | T = inputs.shape[0] |
| 40 | T_lfr = int(np.ceil(T / lfr_n)) |
| 41 | left_padding = inputs[0].repeat((lfr_m - 1) // 2, 1) |
| 42 | inputs = torch.vstack((left_padding, inputs)) |
| 43 | T = T + (lfr_m - 1) // 2 |
| 44 | feat_dim = inputs.shape[-1] |
| 45 | strides = (lfr_n * feat_dim, 1) |
| 46 | sizes = (T_lfr, lfr_m * feat_dim) |
| 47 | last_idx = (T - lfr_m) // lfr_n + 1 |
| 48 | num_padding = lfr_m - (T - last_idx * lfr_n) |
| 49 | if num_padding > 0: |
| 50 | num_padding = ( |
| 51 | (2 * lfr_m - 2 * T + (T_lfr - 1 + last_idx) * lfr_n) |
| 52 | / 2 |
| 53 | * (T_lfr - last_idx) |
| 54 | ) |
| 55 | inputs = torch.vstack([inputs] + [inputs[-1:]] * int(num_padding)) |
| 56 | LFR_outputs = inputs.as_strided(sizes, strides) |
| 57 | return LFR_outputs.clone().type(torch.float32) |
| 58 | |
| 59 | |
| 60 | def load_cmvn(cmvn_file): |
| 61 | with open(cmvn_file, encoding="utf-8") as f: |
| 62 | lines = f.readlines() |
| 63 | means_list = [] |
| 64 | vars_list = [] |
| 65 | for i in range(len(lines)): |
| 66 | line_item = lines[i].split() |
| 67 | if line_item[0] == "<AddShift>": |
| 68 | line_item = lines[i + 1].split() |
| 69 | if line_item[0] == "<LearnRateCoef>": |
| 70 | add_shift_line = line_item[3 : (len(line_item) - 1)] |
| 71 | means_list = list(add_shift_line) |
| 72 | continue |
| 73 | elif line_item[0] == "<Rescale>": |
| 74 | line_item = lines[i + 1].split() |
| 75 | if line_item[0] == "<LearnRateCoef>": |
| 76 | rescale_line = line_item[3 : (len(line_item) - 1)] |
| 77 | vars_list = list(rescale_line) |
| 78 | continue |
| 79 | means = np.array(means_list).astype(np.float32) |
| 80 | vars = np.array(vars_list).astype(np.float32) |
| 81 | cmvn = np.array([means, vars]) |
| 82 | cmvn = torch.as_tensor(cmvn, dtype=torch.float32) |
| 83 | return cmvn |
| 84 | |
| 85 | |
| 86 | class WavFrontend(nn.Module): |
| 87 | """Conventional frontend structure for ASR.""" |
| 88 | |
| 89 | def __init__( |
| 90 | self, |
| 91 | cmvn_file: str = "null", |
| 92 | fs: int = 16000, |
| 93 | window: str = "hamming", |
| 94 | n_mels: int = 80, |
| 95 | frame_length: int = 25, |
| 96 | frame_shift: int = 10, |
| 97 | filter_length_min: int = -1, |
| 98 | filter_length_max: int = -1, |
| 99 | lfr_m: int = 1, |
| 100 | lfr_n: int = 1, |
| 101 | dither: float = 1.0, |
| 102 | snip_edges: bool = True, |
| 103 | upsacle_samples: bool = True, |
| 104 | **kwargs, |
| 105 | ): |
| 106 | super().__init__() |
| 107 | self.fs = fs |
| 108 | self.window = window |
| 109 | self.n_mels = n_mels |
| 110 | self.frame_length = frame_length |
| 111 | self.frame_shift = frame_shift |
| 112 | self.filter_length_min = filter_length_min |
| 113 | self.filter_length_max = filter_length_max |
| 114 | self.lfr_m = lfr_m |
| 115 | self.lfr_n = lfr_n |
| 116 | self.cmvn_file = cmvn_file |
| 117 | self.dither = dither |
| 118 | self.snip_edges = snip_edges |
| 119 | self.upsacle_samples = upsacle_samples |
| 120 | self.cmvn = None if self.cmvn_file is None else load_cmvn(self.cmvn_file) |
| 121 | |
| 122 | def output_size(self) -> int: |
| 123 | return self.n_mels * self.lfr_m |
| 124 | |
| 125 | def forward( |
| 126 | self, |
| 127 | input: torch.Tensor, |
| 128 | input_lengths, |
| 129 | **kwargs, |
| 130 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 131 | batch_size = input.size(0) |
| 132 | feats = [] |
| 133 | feats_lens = [] |
| 134 | for i in range(batch_size): |
| 135 | waveform_length = input_lengths[i] |
| 136 | waveform = input[i][:waveform_length] |
| 137 | if self.upsacle_samples: |
| 138 | waveform = waveform * (1 << 15) |
| 139 | waveform = waveform.unsqueeze(0) |
| 140 | mat = kaldi.fbank( |
| 141 | waveform, |
| 142 | num_mel_bins=self.n_mels, |
| 143 | frame_length=min(self.frame_length, waveform_length / self.fs * 1000), |
| 144 | frame_shift=self.frame_shift, |
| 145 | dither=self.dither, |
| 146 | energy_floor=0.0, |
| 147 | window_type=self.window, |
| 148 | sample_frequency=self.fs, |
| 149 | snip_edges=self.snip_edges, |
| 150 | ) |
| 151 | |
| 152 | if self.lfr_m != 1 or self.lfr_n != 1: |
| 153 | mat = apply_lfr(mat, self.lfr_m, self.lfr_n) |
| 154 | if self.cmvn is not None: |
| 155 | mat = apply_cmvn(mat, self.cmvn) |
| 156 | feat_length = mat.size(0) |
| 157 | feats.append(mat) |
| 158 | feats_lens.append(feat_length) |
| 159 | |
| 160 | feats_lens = torch.as_tensor(feats_lens) |
| 161 | if batch_size == 1: |
| 162 | feats_pad = feats[0][None, :, :] |
| 163 | else: |
| 164 | feats_pad = pad_sequence(feats, batch_first=True, padding_value=0.0) |
| 165 | return feats_pad, feats_lens |
| 166 | |
| 167 | |
| 168 | class FunASRFeatureExtractor(SequenceFeatureExtractor): |
| 169 | r"""Constructs a FunASR feature extractor. |
| 170 | |
| 171 | This feature extractor inherits from [`~feature_extraction_sequence_ |
| 172 | utils.SequenceFeatureExtractor`] which contains most of the main |
| 173 | methods. Users should refer to this superclass for more information |
| 174 | regarding those methods. |
| 175 | |
| 176 | This class extracts mel-filter bank features from raw speech using a custom |
| 177 | numpy implementation of the `Short Time Fourier Transform` which should |
| 178 | match pytorch's `torch.stft` equivalent. |
| 179 | |
| 180 | Args: |
| 181 | feature_size (`int`, *optional*, defaults to 80): |
| 182 | The feature dimension of the extracted features. |
| 183 | sampling_rate (`int`, *optional*, defaults to 16000): |
| 184 | The sampling rate at which the audio files should be digitalized |
| 185 | expressed in hertz (Hz). |
| 186 | hop_length (`int`, *optional*, defaults to 160): |
| 187 | Length of the overlapping windows for the STFT used to obtain the |
| 188 | Mel Frequency coefficients. |
| 189 | chunk_length (`int`, *optional*, defaults to 30): |
| 190 | The maximum number of chunks of `sampling_rate` samples used to |
| 191 | trim and pad longer or shorter audio sequences. |
| 192 | n_fft (`int`, *optional*, defaults to 400): |
| 193 | Size of the Fourier transform. |
| 194 | padding_value (`float`, *optional*, defaults to 0.0): |
| 195 | Padding value used to pad the audio. Should correspond to silences. |
| 196 | dither (`float`, *optional*, defaults to 0.0): |
| 197 | Adds dithering. In other words, adds a small Gaussian noise to each frame. |
| 198 | E.g. use 0.0001 to add dithering with a normal distribution centered |
| 199 | around 0.0 with standard deviation 0.0001 (assuming [-1,+1] range |
| 200 | of raw_speech). The value 0.0 means no dithering. |
| 201 | Dithering has similar effect as `spectrogram(mel_floor=...)`. It reduces |
| 202 | the high log_mel_fbank values for signals with hard-zero sections, |
| 203 | when VAD cutoff is present in the signal. |
| 204 | |
| 205 | """ |
| 206 | |
| 207 | model_input_names = ["input_features"] |
| 208 | |
| 209 | def __init__( |
| 210 | self, |
| 211 | feature_size=80, |
| 212 | sampling_rate=16000, |
| 213 | hop_length=160, |
| 214 | chunk_length=30, |
| 215 | n_fft=400, |
| 216 | padding_value=0.0, |
| 217 | dither=0.0, |
| 218 | max_length=1000, |
| 219 | return_attention_mask=False, |
| 220 | **kwargs, |
| 221 | ): |
| 222 | super().__init__( |
| 223 | feature_size=feature_size, |
| 224 | sampling_rate=sampling_rate, |
| 225 | padding_value=padding_value, |
| 226 | return_attention_mask=return_attention_mask, |
| 227 | **kwargs, |
| 228 | ) |
| 229 | self.frontend_conf = kwargs.get("frontend_conf", {}) |
| 230 | self.max_length = max_length |
| 231 | self.n_fft = n_fft |
| 232 | self.hop_length = hop_length |
| 233 | self.chunk_length = chunk_length |
| 234 | self.n_samples = chunk_length * sampling_rate |
| 235 | self.nb_max_frames = self.n_samples // hop_length |
| 236 | self.sampling_rate = sampling_rate |
| 237 | self.dither = dither |
| 238 | |
| 239 | def extract_fbank( |
| 240 | self, data, data_len=None, data_type: str = "sound", frontend=None, **kwargs |
| 241 | ): |
| 242 | if isinstance(data, np.ndarray): |
| 243 | data = torch.from_numpy(data) |
| 244 | if len(data.shape) < 2: |
| 245 | data = data[None, :] # data: [batch, N] |
| 246 | data_len = [data.shape[1]] if data_len is None else data_len |
| 247 | elif isinstance(data, torch.Tensor): |
| 248 | if len(data.shape) < 2: |
| 249 | data = data[None, :] # data: [batch, N] |
| 250 | data_len = [data.shape[1]] if data_len is None else data_len |
| 251 | elif isinstance(data, (list, tuple)): |
| 252 | data_list, data_len = [], [] |
| 253 | for data_i in data: |
| 254 | if isinstance(data_i, np.ndarray): |
| 255 | data_i = torch.from_numpy(data_i) |
| 256 | data_list.append(data_i) |
| 257 | data_len.append(data_i.shape[0]) |
| 258 | data = pad_sequence(data_list, batch_first=True) |
| 259 | |
| 260 | data, data_len = frontend(data, data_len, **kwargs) |
| 261 | |
| 262 | if isinstance(data_len, (list, tuple)): |
| 263 | data_len = torch.tensor([data_len]) |
| 264 | return data.to(torch.float32), data_len.to(torch.int32) |
| 265 | |
| 266 | def __call__( |
| 267 | self, |
| 268 | raw_speech: np.ndarray | list[float] | list[np.ndarray] | list[list[float]], |
| 269 | truncation: bool = True, |
| 270 | pad_to_multiple_of: int | None = None, |
| 271 | return_tensors: str | TensorType | None = None, |
| 272 | return_attention_mask: bool | None = None, |
| 273 | padding: str | None = "max_length", |
| 274 | max_length: int | None = None, |
| 275 | sampling_rate: int | None = None, |
| 276 | do_normalize: bool | None = None, |
| 277 | device: str | None = "cpu", |
| 278 | return_token_timestamps: bool | None = None, |
| 279 | **kwargs, |
| 280 | ) -> BatchFeature: |
| 281 | frontend = WavFrontend(**self.frontend_conf, dither=self.dither) |
| 282 | |
| 283 | feats = [] |
| 284 | speech_lengths = [] |
| 285 | fake_token_lengths = [] |
| 286 | for speech in raw_speech: |
| 287 | feature, length = self.extract_fbank( |
| 288 | speech, |
| 289 | data_type=kwargs.get("data_type", "sound"), |
| 290 | frontend=frontend, |
| 291 | is_final=True, |
| 292 | ) |
| 293 | feats.append(feature) |
| 294 | speech_lengths.append(length) |
| 295 | olens = 1 + (length - 3 + 2 * 1) // 2 |
| 296 | olens = 1 + (olens - 3 + 2 * 1) // 2 |
| 297 | fake_token_len = (olens - 1) // 2 + 1 |
| 298 | fake_token_len = torch.clamp(fake_token_len, min=1) |
| 299 | fake_token_lengths.append(fake_token_len) |
| 300 | |
| 301 | feats = torch.concat(feats, dim=0) |
| 302 | batched_speech = self.pad( |
| 303 | BatchFeature({"input_features": feats}), |
| 304 | padding=padding, |
| 305 | max_length=max_length if max_length else self.max_length, |
| 306 | truncation=truncation, |
| 307 | pad_to_multiple_of=pad_to_multiple_of, |
| 308 | return_attention_mask=return_attention_mask or do_normalize, |
| 309 | ) |
| 310 | if return_tensors is not None: |
| 311 | batched_speech = batched_speech.convert_to_tensors(return_tensors) |
| 312 | |
| 313 | batched_speech["speech_lengths"] = torch.tensor(speech_lengths) |
| 314 | batched_speech["fake_token_lengths"] = torch.concat(fake_token_lengths) |
| 315 | return batched_speech |
| 316 | |
| 317 | |
| 318 | class FunASRProcessor(ProcessorMixin): |
| 319 | r"""Constructs a FunASR processor which wraps a FunASR feature extractor and |
| 320 | a FunASR tokenizer into a single processor. |
| 321 | |
| 322 | [`FunASRProcessor`] offers all the functionalities of |
| 323 | [`FunASRFeatureExtractor`] and [`Qwen2Tokenizer`]. See the |
| 324 | [`~FunASRProcessor.__call__`] and [`~FunASRProcessor.decode`] for more |
| 325 | information. |
| 326 | |
| 327 | Args: |
| 328 | feature_extractor (`FunASRFeatureExtractor`): An instance of |
| 329 | [`FunASRFeatureExtractor`]. |
| 330 | The feature extractor is a required input. |
| 331 | tokenizer (`Qwen2Tokenizer`): |
| 332 | An instance of [`Qwen2Tokenizer`]. The tokenizer is a required |
| 333 | input. |
| 334 | |
| 335 | """ |
| 336 | |
| 337 | feature_extractor_class = "FunASRFeatureExtractor" |
| 338 | tokenizer_class = ("Qwen2Tokenizer", "Qwen2TokenizerFast") |
| 339 | |
| 340 | def __init__( |
| 341 | self, |
| 342 | feature_extractor, |
| 343 | tokenizer, |
| 344 | audio_token="<|AUDIO|>", |
| 345 | ): |
| 346 | super().__init__(feature_extractor, tokenizer) |
| 347 | self.current_processor = self.feature_extractor |
| 348 | self._in_target_context_manager = False |
| 349 | self.audio_token = ( |
| 350 | tokenizer.audio_token if hasattr(tokenizer, "audio_token") else audio_token |
| 351 | ) |
| 352 | self.audio_token_id = tokenizer.convert_tokens_to_ids(self.audio_token) |
| 353 | |
| 354 | def get_decoder_prompt_ids(self, task=None, language=None, no_timestamps=True): |
| 355 | return self.tokenizer.get_decoder_prompt_ids( |
| 356 | task=task, language=language, no_timestamps=no_timestamps |
| 357 | ) |
| 358 | |
| 359 | def __call__(self, *args, **kwargs): |
| 360 | """Forwards the `audio` argument to FunASRFeatureExtractor's |
| 361 | [`~FunASRFeatureExtractor.__call__`] and the `text` argument to |
| 362 | [`~Qwen2Tokenizer.__call__`]. Please refer to the docstring of the |
| 363 | above two methods for more information. |
| 364 | """ |
| 365 | if self._in_target_context_manager: |
| 366 | return self.current_processor(*args, **kwargs) |
| 367 | |
| 368 | audio = kwargs.pop("audio", None) |
| 369 | sampling_rate = kwargs.pop("sampling_rate", None) |
| 370 | text = kwargs.pop("text", None) |
| 371 | if len(args) > 0: |
| 372 | audio = args[0] |
| 373 | args = args[1:] |
| 374 | |
| 375 | if text is None: |
| 376 | raise ValueError("You need to specify `text` input to process.") |
| 377 | elif isinstance(text, str): |
| 378 | text = [text] |
| 379 | elif not isinstance(text, list) and not isinstance(text[0], str): |
| 380 | raise ValueError( |
| 381 | "Invalid input text. Please provide a string, or a list of strings" |
| 382 | ) |
| 383 | |
| 384 | if audio is not None: |
| 385 | # ensure we have as much audios as audio tokens |
| 386 | num_audio_tokens = sum(sample.count(self.audio_token) for sample in text) |
| 387 | num_audios = 1 if type(audio) is np.ndarray else len(audio) |
| 388 | if num_audio_tokens != num_audios: |
| 389 | raise ValueError( |
| 390 | f"Found {num_audio_tokens} {self.audio_token} token{'s' if num_audio_tokens > 1 else ''} in provided text but received {num_audios} audio{'s' if num_audios > 1 else ''}" # noqa: E501 |
| 391 | ) |
| 392 | inputs = self.feature_extractor( |
| 393 | audio, *args, sampling_rate=sampling_rate, **kwargs |
| 394 | ) |
| 395 | |
| 396 | expanded_text = [] |
| 397 | for sample in text: |
| 398 | replace_str = [] |
| 399 | while self.audio_token in sample: |
| 400 | num_audio_tokens = inputs["fake_token_lengths"].item() |
| 401 | |
| 402 | expanded_audio_token = self.audio_token * num_audio_tokens |
| 403 | |
| 404 | replace_str.append(expanded_audio_token) |
| 405 | sample = sample.replace(self.audio_token, "<placeholder>", 1) |
| 406 | |
| 407 | while "<placeholder>" in sample: |
| 408 | sample = sample.replace("<placeholder>", replace_str.pop(0), 1) |
| 409 | expanded_text.append(sample) |
| 410 | text = expanded_text |
| 411 | |
| 412 | if text is not None: |
| 413 | encodings = self.tokenizer(text, **kwargs) |
| 414 | |
| 415 | if text is None: |
| 416 | return inputs |
| 417 | |
| 418 | elif audio is None: |
| 419 | return encodings |
| 420 | else: |
| 421 | inputs["labels"] = encodings["input_ids"] |
| 422 | |
| 423 | return inputs |
| 424 | |
| 425 | def get_prompt_ids(self, text: str, return_tensors="np"): |
| 426 | return self.tokenizer.get_prompt_ids(text, return_tensors=return_tensors) |
| 427 | |
| 428 | |
| 429 | AutoFeatureExtractor.register("FunASRFeatureExtractor", FunASRFeatureExtractor) |