Spaces:
Sleeping
Sleeping
File size: 13,547 Bytes
e11256b |
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 |
"""
Scraibe Class
--------------------
This class serves as the core of the transcription system, responsible for handling
transcription and diarization of audio files. It leverages pretrained models for
speech-to-text (such as Whisper) and speaker diarization (such as pyannote.audio),
providing an accessible interface for audio processing tasks such as transcription,
speaker separation, and timestamping.
By encapsulating the complexities of underlying models, it allows for straightforward
integration into various applications, ranging from transcription services to voice assistants.
Available Classes:
- Scraibe: Main class for performing transcription and diarization.
Includes methods for loading models, processing audio files,
and formatting the transcription output.
Usage:
from scraibe import Scraibe
model = Scraibe()
transcript = model.autotranscribe("path/to/audiofile.wav")
"""
# Standard Library Imports
import os
from glob import iglob
from subprocess import run
from typing import TypeVar, Union
from warnings import warn
# Third-Party Imports
import torch
from numpy import ndarray
from tqdm import trange
# Application-Specific Imports
from .audio import AudioProcessor
from .diarisation import Diariser
from .transcriber import Transcriber, load_transcriber, whisper
from .transcript_exporter import Transcript
from .misc import SCRAIBE_TORCH_DEVICE
DiarisationType = TypeVar('DiarisationType')
class Scraibe:
"""
Scraibe is a class responsible for managing the transcription and diarization of audio files.
It serves as the core of the transcription system, incorporating pretrained models
for speech-to-text (such as Whisper) and speaker diarization (such as pyannote.audio),
allowing for comprehensive audio processing.
Attributes:
transcriber (Transcriber): The transcriber object to handle transcription.
diariser (Diariser): The diariser object to handle diarization.
Methods:
__init__: Initializes the Scraibe class with appropriate models.
transcribe: Transcribes an audio file using the whisper model and pyannote diarization model.
remove_audio_file: Removes the original audio file to avoid disk space issues or ensure data privacy.
get_audio_file: Gets an audio file as an AudioProcessor object.
"""
def __init__(self,
whisper_model: Union[bool, str, whisper] = None,
whisper_type: str = "whisper",
dia_model: Union[bool, str, DiarisationType] = None,
**kwargs) -> None:
"""Initializes the Scraibe class.
Args:
whisper_model (Union[bool, str, whisper], optional):
Path to whisper model or whisper model itself.
whisper_type (str):
Type of whisper model to load. "whisper" or "faster-whisper".
diarisation_model (Union[bool, str, DiarisationType], optional):
Path to pyannote diarization model or model itself.
**kwargs: Additional keyword arguments for whisper
and pyannote diarization models.
e.g.:
- verbose: If True, the class will print additional information.
- save_kwargs: If True, the keyword arguments will be saved
for autotranscribe. So you can unload the class and reload it again.
"""
if whisper_model is None:
self.transcriber = load_transcriber(
"large-v3", whisper_type, **kwargs)
elif isinstance(whisper_model, str):
self.transcriber = load_transcriber(
whisper_model, whisper_type, **kwargs)
else:
self.transcriber = whisper_model
if dia_model is None:
self.diariser = Diariser.load_model(**kwargs)
elif isinstance(dia_model, str):
self.diariser = Diariser.load_model(dia_model, **kwargs)
else:
self.diariser: Diariser = dia_model
if kwargs.get("verbose"):
print("Scraibe initialized all models successfully loaded.")
self.verbose = True
else:
self.verbose = False
# Save kwargs for autotranscribe if you want to unload the class and load it again.
if kwargs.get('save_setup'):
self.params = dict(whisper_model=whisper_model,
dia_model=dia_model,
**kwargs)
else:
self.params = {}
self.device = kwargs.get(
"device", SCRAIBE_TORCH_DEVICE)
def autotranscribe(self, audio_file: Union[str, torch.Tensor, ndarray],
remove_original: bool = False,
**kwargs) -> Transcript:
"""
Transcribes an audio file using the whisper model and pyannote diarization model.
Args:
audio_file (Union[str, torch.Tensor, ndarray]):
Path to audio file or a tensor representing the audio.
remove_original (bool, optional): If True, the original audio file will
be removed after transcription.
*args: Additional positional arguments for diarization and transcription.
**kwargs: Additional keyword arguments for diarization and transcription.
Returns:
Transcript: A Transcript object containing the transcription,
which can be exported to different formats.
"""
if kwargs.get("verbose"):
self.verbose = kwargs.get("verbose")
# Get audio file as an AudioProcessor object
audio_file: AudioProcessor = self.get_audio_file(audio_file)
# Prepare waveform and sample rate for diarization
dia_audio = {
"waveform": audio_file.waveform.reshape(1, len(audio_file.waveform)).to(self.device),
"sample_rate": audio_file.sr
}
if self.verbose:
print("Starting diarisation.")
diarisation = self.diariser.diarization(dia_audio, **kwargs)
if not diarisation["segments"]:
print("No segments found. Try to run transcription without diarisation.")
transcript = self.transcriber.transcribe(
audio_file.waveform, **kwargs)
final_transcript = {0: {"speakers": 'SPEAKER_01',
"segments": [0, len(audio_file.waveform)],
"text": transcript}}
return Transcript(final_transcript)
if self.verbose:
print("Diarisation finished. Starting transcription.")
# Transcribe each segment and store the results
final_transcript = dict()
for i in trange(len(diarisation["segments"]), desc="Transcribing", disable=not self.verbose):
seg = diarisation["segments"][i]
audio = audio_file.cut(seg[0], seg[1])
transcript = self.transcriber.transcribe(audio, **kwargs)
final_transcript[i] = {"speakers": diarisation["speakers"][i],
"segments": seg,
"text": transcript}
# Remove original file if needed
if remove_original:
if kwargs.get("shred") is True:
self.remove_audio_file(audio_file, shred=True)
else:
self.remove_audio_file(audio_file, shred=False)
return Transcript(final_transcript)
def diarization(self, audio_file: Union[str, torch.Tensor, ndarray],
**kwargs) -> dict:
"""
Perform diarization on an audio file using the pyannote diarization model.
Args:
audio_file (Union[str, torch.Tensor, ndarray]):
The audio source which can either be a path to the audio file or a tensor representation.
**kwargs:
Additional keyword arguments for diarization.
Returns:
dict:
A dictionary containing the results of the diarization process.
"""
# Get audio file as an AudioProcessor object
audio_file: AudioProcessor = self.get_audio_file(audio_file)
# Prepare waveform and sample rate for diarization
dia_audio = {
"waveform": audio_file.waveform.reshape(1, len(audio_file.waveform)).to(self.device),
"sample_rate": audio_file.sr
}
print("Starting diarisation.")
diarisation = self.diariser.diarization(dia_audio, **kwargs)
return diarisation
def transcribe(self, audio_file: Union[str, torch.Tensor, ndarray],
**kwargs):
"""
Transcribe the provided audio file.
Args:
audio_file (Union[str, torch.Tensor, ndarray]):
The audio source, which can either be a path or a tensor representation.
**kwargs:
Additional keyword arguments for transcription.
Returns:
str:
The transcribed text from the audio source.
"""
audio_file: AudioProcessor = self.get_audio_file(audio_file)
return self.transcriber.transcribe(audio_file.waveform, **kwargs)
def update_transcriber(self, whisper_model: Union[str, whisper], **kwargs) -> None:
"""
Update the transcriber model.
Args:
whisper_model (Union[str, whisper]):
The new whisper model to use for transcription.
**kwargs:
Additional keyword arguments for the transcriber model.
Returns:
None
"""
_old_model = self.transcriber.model_name
if isinstance(whisper_model, str):
self.transcriber = load_transcriber(whisper_model, **kwargs)
elif isinstance(whisper_model, Transcriber):
self.transcriber = whisper_model
else:
warn(
f"Invalid model type. Please provide a valid model. Fallback to old {_old_model} Model.", RuntimeWarning)
return None
def update_diariser(self, dia_model: Union[str, DiarisationType], **kwargs) -> None:
"""
Update the diariser model.
Args:
dia_model (Union[str, DiarisationType]):
The new diariser model to use for diarization.
**kwargs:
Additional keyword arguments for the diariser model.
Returns:
None
"""
if isinstance(dia_model, str):
self.diariser = Diariser.load_model(dia_model, **kwargs)
elif isinstance(dia_model, Diariser):
self.diariser = dia_model
else:
warn("Invalid model type. Please provide a valid model. Fallback to old Model.", RuntimeWarning)
return None
@staticmethod
def remove_audio_file(audio_file: str,
shred: bool = False) -> None:
"""
Removes the original audio file to avoid disk space issues or ensure data privacy.
Args:
audio_file_path (str): Path to the audio file.
shred (bool, optional): If True, the audio file will be shredded,
not just removed.
"""
if not os.path.exists(audio_file):
raise ValueError(f"Audiofile {audio_file} does not exist.")
if shred:
warn("Shredding audiofile can take a long time.", RuntimeWarning)
gen = iglob(f'{audio_file}', recursive=True)
cmd = ['shred', '-zvu', '-n', '10', f'{audio_file}']
if os.path.isdir(audio_file):
raise ValueError(f"Audiofile {audio_file} is a directory.")
for file in gen:
print(f'shredding {file} now\n')
run(cmd, check=True)
else:
os.remove(audio_file)
print(f"Audiofile {audio_file} removed.")
@staticmethod
def get_audio_file(audio_file: Union[str, torch.Tensor, ndarray]) -> AudioProcessor:
"""Gets an audio file as TorchAudioProcessor.
Args:
audio_file (Union[str, torch.Tensor, ndarray]): Path to the audio file or
a tensor representing the audio.
*args: Additional positional arguments.
**kwargs: Additional keyword arguments.
Returns:
AudioProcessor: An object containing the waveform and sample rate in
torch.Tensor format.
"""
if isinstance(audio_file, str):
audio_file = AudioProcessor.from_file(audio_file)
elif isinstance(audio_file, torch.Tensor):
audio_file = AudioProcessor(audio_file[0], audio_file[1])
elif isinstance(audio_file, ndarray):
audio_file = AudioProcessor(torch.Tensor(audio_file[0]),
audio_file[1])
if not isinstance(audio_file, AudioProcessor):
raise ValueError(f'Audiofile must be of type AudioProcessor,'
f'not {type(audio_file)}')
return audio_file
def __repr__(self):
return f"Scraibe(transcriber={self.transcriber}, diariser={self.diariser})"
|