Spaces:
Sleeping
Sleeping
add deepspeed
Browse files- inference.py +319 -0
- requirements.txt +2 -1
- src/AI_Models/wave2vec_inference.py +560 -264
- src/apis/controllers/speaking_controller.py +7 -7
- src/apis/routes/speaking_route.py +3 -0
inference.py
ADDED
|
@@ -0,0 +1,319 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from transformers import (
|
| 3 |
+
Wav2Vec2ForCTC,
|
| 4 |
+
Wav2Vec2Processor,
|
| 5 |
+
AutoProcessor,
|
| 6 |
+
AutoModelForCTC,
|
| 7 |
+
)
|
| 8 |
+
|
| 9 |
+
# import deepspeed
|
| 10 |
+
import librosa
|
| 11 |
+
import numpy as np
|
| 12 |
+
from typing import Optional, List, Union
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
def get_model_name(model_name: Optional[str] = None) -> str:
|
| 16 |
+
"""Helper function to get model name with default fallback"""
|
| 17 |
+
if model_name is None:
|
| 18 |
+
return "facebook/wav2vec2-large-robust-ft-libri-960h"
|
| 19 |
+
return model_name
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
class Wave2Vec2Inference:
|
| 23 |
+
def __init__(
|
| 24 |
+
self,
|
| 25 |
+
model_name: Optional[str] = None,
|
| 26 |
+
use_gpu: bool = True,
|
| 27 |
+
use_deepspeed: bool = True,
|
| 28 |
+
):
|
| 29 |
+
"""
|
| 30 |
+
Initialize Wav2Vec2 model for inference with optional DeepSpeed optimization.
|
| 31 |
+
|
| 32 |
+
Args:
|
| 33 |
+
model_name: HuggingFace model name or None for default
|
| 34 |
+
use_gpu: Whether to use GPU acceleration
|
| 35 |
+
use_deepspeed: Whether to use DeepSpeed optimization
|
| 36 |
+
"""
|
| 37 |
+
# Get the actual model name using helper function
|
| 38 |
+
self.model_name = get_model_name(model_name)
|
| 39 |
+
self.use_deepspeed = use_deepspeed
|
| 40 |
+
|
| 41 |
+
# Auto-detect device
|
| 42 |
+
if use_gpu:
|
| 43 |
+
if torch.backends.mps.is_available():
|
| 44 |
+
self.device = "mps"
|
| 45 |
+
elif torch.cuda.is_available():
|
| 46 |
+
self.device = "cuda"
|
| 47 |
+
else:
|
| 48 |
+
self.device = "cpu"
|
| 49 |
+
else:
|
| 50 |
+
self.device = "cpu"
|
| 51 |
+
|
| 52 |
+
print(f"Using device: {self.device}")
|
| 53 |
+
print(f"Loading model: {self.model_name}")
|
| 54 |
+
print(f"DeepSpeed enabled: {self.use_deepspeed}")
|
| 55 |
+
|
| 56 |
+
# Check if model is XLSR and use appropriate processor/model
|
| 57 |
+
is_xlsr = "xlsr" in self.model_name.lower()
|
| 58 |
+
|
| 59 |
+
if is_xlsr:
|
| 60 |
+
print("Using Wav2Vec2Processor and Wav2Vec2ForCTC for XLSR model")
|
| 61 |
+
self.processor = Wav2Vec2Processor.from_pretrained(self.model_name)
|
| 62 |
+
self.model = Wav2Vec2ForCTC.from_pretrained(self.model_name)
|
| 63 |
+
else:
|
| 64 |
+
print("Using AutoProcessor and AutoModelForCTC")
|
| 65 |
+
self.processor = AutoProcessor.from_pretrained(self.model_name)
|
| 66 |
+
self.model = AutoModelForCTC.from_pretrained(self.model_name)
|
| 67 |
+
|
| 68 |
+
# Initialize DeepSpeed if enabled
|
| 69 |
+
if self.use_deepspeed:
|
| 70 |
+
self._init_deepspeed()
|
| 71 |
+
else:
|
| 72 |
+
self.model.to(self.device)
|
| 73 |
+
self.model.eval()
|
| 74 |
+
self.ds_engine = None
|
| 75 |
+
|
| 76 |
+
# Disable gradients for inference
|
| 77 |
+
torch.set_grad_enabled(False)
|
| 78 |
+
|
| 79 |
+
def _init_deepspeed(self):
|
| 80 |
+
"""Initialize DeepSpeed inference engine"""
|
| 81 |
+
try:
|
| 82 |
+
# DeepSpeed configuration based on device
|
| 83 |
+
if self.device == "cuda":
|
| 84 |
+
ds_config = {
|
| 85 |
+
"tensor_parallel": {"tp_size": 1},
|
| 86 |
+
"dtype": torch.float32,
|
| 87 |
+
"replace_with_kernel_inject": True,
|
| 88 |
+
"enable_cuda_graph": False,
|
| 89 |
+
}
|
| 90 |
+
else:
|
| 91 |
+
ds_config = {
|
| 92 |
+
"tensor_parallel": {"tp_size": 1},
|
| 93 |
+
"dtype": torch.float32,
|
| 94 |
+
"replace_with_kernel_inject": False,
|
| 95 |
+
"enable_cuda_graph": False,
|
| 96 |
+
}
|
| 97 |
+
|
| 98 |
+
print("Initializing DeepSpeed inference engine...")
|
| 99 |
+
self.ds_engine = deepspeed.init_inference(self.model, **ds_config)
|
| 100 |
+
self.ds_engine.module.to(self.device)
|
| 101 |
+
|
| 102 |
+
except Exception as e:
|
| 103 |
+
print(f"DeepSpeed initialization failed: {e}")
|
| 104 |
+
print("Falling back to standard PyTorch inference...")
|
| 105 |
+
self.use_deepspeed = False
|
| 106 |
+
self.ds_engine = None
|
| 107 |
+
self.model.to(self.device)
|
| 108 |
+
self.model.eval()
|
| 109 |
+
|
| 110 |
+
def _get_model(self):
|
| 111 |
+
"""Get the appropriate model for inference"""
|
| 112 |
+
if self.use_deepspeed and self.ds_engine is not None:
|
| 113 |
+
return self.ds_engine.module
|
| 114 |
+
return self.model
|
| 115 |
+
|
| 116 |
+
def buffer_to_text(
|
| 117 |
+
self, audio_buffer: Union[np.ndarray, torch.Tensor, List]
|
| 118 |
+
) -> str:
|
| 119 |
+
"""
|
| 120 |
+
Convert audio buffer to text transcription.
|
| 121 |
+
|
| 122 |
+
Args:
|
| 123 |
+
audio_buffer: Audio data as numpy array, tensor, or list
|
| 124 |
+
|
| 125 |
+
Returns:
|
| 126 |
+
str: Transcribed text
|
| 127 |
+
"""
|
| 128 |
+
if len(audio_buffer) == 0:
|
| 129 |
+
return ""
|
| 130 |
+
|
| 131 |
+
# Convert to tensor
|
| 132 |
+
if isinstance(audio_buffer, np.ndarray):
|
| 133 |
+
audio_tensor = torch.from_numpy(audio_buffer).float()
|
| 134 |
+
elif isinstance(audio_buffer, list):
|
| 135 |
+
audio_tensor = torch.tensor(audio_buffer, dtype=torch.float32)
|
| 136 |
+
else:
|
| 137 |
+
audio_tensor = audio_buffer.float()
|
| 138 |
+
|
| 139 |
+
# Process audio
|
| 140 |
+
inputs = self.processor(
|
| 141 |
+
audio_tensor,
|
| 142 |
+
sampling_rate=16_000,
|
| 143 |
+
return_tensors="pt",
|
| 144 |
+
padding=True,
|
| 145 |
+
)
|
| 146 |
+
|
| 147 |
+
# Move to device
|
| 148 |
+
input_values = inputs.input_values.to(self.device)
|
| 149 |
+
attention_mask = (
|
| 150 |
+
inputs.attention_mask.to(self.device)
|
| 151 |
+
if "attention_mask" in inputs
|
| 152 |
+
else None
|
| 153 |
+
)
|
| 154 |
+
|
| 155 |
+
# Get the appropriate model
|
| 156 |
+
model = self._get_model()
|
| 157 |
+
|
| 158 |
+
# Inference
|
| 159 |
+
with torch.no_grad():
|
| 160 |
+
if attention_mask is not None:
|
| 161 |
+
outputs = model(input_values, attention_mask=attention_mask)
|
| 162 |
+
else:
|
| 163 |
+
outputs = model(input_values)
|
| 164 |
+
|
| 165 |
+
# Handle different output formats
|
| 166 |
+
if hasattr(outputs, "logits"):
|
| 167 |
+
logits = outputs.logits
|
| 168 |
+
else:
|
| 169 |
+
logits = outputs
|
| 170 |
+
|
| 171 |
+
# Decode
|
| 172 |
+
predicted_ids = torch.argmax(logits, dim=-1)
|
| 173 |
+
if self.device != "cpu":
|
| 174 |
+
predicted_ids = predicted_ids.cpu()
|
| 175 |
+
|
| 176 |
+
transcription = self.processor.batch_decode(predicted_ids)[0]
|
| 177 |
+
return transcription.lower().strip()
|
| 178 |
+
|
| 179 |
+
def file_to_text(self, filename: str) -> str:
|
| 180 |
+
"""
|
| 181 |
+
Transcribe audio file to text.
|
| 182 |
+
|
| 183 |
+
Args:
|
| 184 |
+
filename: Path to audio file
|
| 185 |
+
|
| 186 |
+
Returns:
|
| 187 |
+
str: Transcribed text
|
| 188 |
+
"""
|
| 189 |
+
try:
|
| 190 |
+
audio_input, _ = librosa.load(filename, sr=16000, dtype=np.float32)
|
| 191 |
+
return self.buffer_to_text(audio_input)
|
| 192 |
+
except Exception as e:
|
| 193 |
+
print(f"Error loading audio file {filename}: {e}")
|
| 194 |
+
return ""
|
| 195 |
+
|
| 196 |
+
def batch_file_to_text(self, filenames: List[str]) -> List[str]:
|
| 197 |
+
"""
|
| 198 |
+
Transcribe multiple audio files to text.
|
| 199 |
+
|
| 200 |
+
Args:
|
| 201 |
+
filenames: List of audio file paths
|
| 202 |
+
|
| 203 |
+
Returns:
|
| 204 |
+
List[str]: List of transcribed texts
|
| 205 |
+
"""
|
| 206 |
+
results = []
|
| 207 |
+
for i, filename in enumerate(filenames):
|
| 208 |
+
print(f"Processing file {i+1}/{len(filenames)}: {filename}")
|
| 209 |
+
transcription = self.file_to_text(filename)
|
| 210 |
+
results.append(transcription)
|
| 211 |
+
if transcription:
|
| 212 |
+
print(f"Transcription: {transcription}")
|
| 213 |
+
else:
|
| 214 |
+
print("Failed to transcribe")
|
| 215 |
+
return results
|
| 216 |
+
|
| 217 |
+
def transcribe_with_confidence(
|
| 218 |
+
self, audio_buffer: Union[np.ndarray, torch.Tensor]
|
| 219 |
+
) -> tuple:
|
| 220 |
+
"""
|
| 221 |
+
Transcribe audio and return confidence scores.
|
| 222 |
+
|
| 223 |
+
Args:
|
| 224 |
+
audio_buffer: Audio data
|
| 225 |
+
|
| 226 |
+
Returns:
|
| 227 |
+
tuple: (transcription, confidence_scores)
|
| 228 |
+
"""
|
| 229 |
+
if len(audio_buffer) == 0:
|
| 230 |
+
return "", []
|
| 231 |
+
|
| 232 |
+
# Convert to tensor
|
| 233 |
+
if isinstance(audio_buffer, np.ndarray):
|
| 234 |
+
audio_tensor = torch.from_numpy(audio_buffer).float()
|
| 235 |
+
else:
|
| 236 |
+
audio_tensor = audio_buffer.float()
|
| 237 |
+
|
| 238 |
+
# Process audio
|
| 239 |
+
inputs = self.processor(
|
| 240 |
+
audio_tensor,
|
| 241 |
+
sampling_rate=16_000,
|
| 242 |
+
return_tensors="pt",
|
| 243 |
+
padding=True,
|
| 244 |
+
)
|
| 245 |
+
|
| 246 |
+
input_values = inputs.input_values.to(self.device)
|
| 247 |
+
attention_mask = (
|
| 248 |
+
inputs.attention_mask.to(self.device)
|
| 249 |
+
if "attention_mask" in inputs
|
| 250 |
+
else None
|
| 251 |
+
)
|
| 252 |
+
|
| 253 |
+
model = self._get_model()
|
| 254 |
+
|
| 255 |
+
# Inference
|
| 256 |
+
with torch.no_grad():
|
| 257 |
+
if attention_mask is not None:
|
| 258 |
+
outputs = model(input_values, attention_mask=attention_mask)
|
| 259 |
+
else:
|
| 260 |
+
outputs = model(input_values)
|
| 261 |
+
|
| 262 |
+
if hasattr(outputs, "logits"):
|
| 263 |
+
logits = outputs.logits
|
| 264 |
+
else:
|
| 265 |
+
logits = outputs
|
| 266 |
+
|
| 267 |
+
# Get probabilities and confidence scores
|
| 268 |
+
probs = torch.nn.functional.softmax(logits, dim=-1)
|
| 269 |
+
predicted_ids = torch.argmax(logits, dim=-1)
|
| 270 |
+
|
| 271 |
+
# Calculate confidence as max probability for each prediction
|
| 272 |
+
max_probs = torch.max(probs, dim=-1)[0]
|
| 273 |
+
confidence_scores = max_probs.cpu().numpy().tolist()
|
| 274 |
+
|
| 275 |
+
if self.device != "cpu":
|
| 276 |
+
predicted_ids = predicted_ids.cpu()
|
| 277 |
+
|
| 278 |
+
transcription = self.processor.batch_decode(predicted_ids)[0]
|
| 279 |
+
return transcription.lower().strip(), confidence_scores
|
| 280 |
+
|
| 281 |
+
def cleanup(self):
|
| 282 |
+
"""Clean up resources"""
|
| 283 |
+
if hasattr(self, "ds_engine") and self.ds_engine is not None:
|
| 284 |
+
del self.ds_engine
|
| 285 |
+
if hasattr(self, "model"):
|
| 286 |
+
del self.model
|
| 287 |
+
if hasattr(self, "processor"):
|
| 288 |
+
del self.processor
|
| 289 |
+
torch.cuda.empty_cache() if torch.cuda.is_available() else None
|
| 290 |
+
|
| 291 |
+
def __del__(self):
|
| 292 |
+
"""Destructor to clean up resources"""
|
| 293 |
+
self.cleanup()
|
| 294 |
+
|
| 295 |
+
|
| 296 |
+
# Example usage
|
| 297 |
+
if __name__ == "__main__":
|
| 298 |
+
# Initialize with DeepSpeed
|
| 299 |
+
asr = Wave2Vec2Inference(
|
| 300 |
+
model_name="facebook/wav2vec2-large-robust-ft-libri-960h",
|
| 301 |
+
use_gpu=False,
|
| 302 |
+
use_deepspeed=False,
|
| 303 |
+
)
|
| 304 |
+
|
| 305 |
+
# Single file transcription
|
| 306 |
+
result = asr.file_to_text("./test_audio/hello_how_are_you_today.wav")
|
| 307 |
+
print(f"Transcription: {result}")
|
| 308 |
+
|
| 309 |
+
# # Batch processing
|
| 310 |
+
# files = ["audio1.wav", "audio2.wav", "audio3.wav"]
|
| 311 |
+
# batch_results = asr.batch_file_to_text(files)
|
| 312 |
+
|
| 313 |
+
# # Transcription with confidence scores
|
| 314 |
+
# audio_data, _ = librosa.load("path/to/audio.wav", sr=16000)
|
| 315 |
+
# transcription, confidence = asr.transcribe_with_confidence(audio_data)
|
| 316 |
+
# print(f"Transcription: {transcription}")
|
| 317 |
+
# print(f"Average confidence: {np.mean(confidence):.3f}")
|
| 318 |
+
|
| 319 |
+
# Cleanup
|
requirements.txt
CHANGED
|
@@ -23,4 +23,5 @@ onnx
|
|
| 23 |
transformers
|
| 24 |
torch
|
| 25 |
optimum[onnxruntime]
|
| 26 |
-
Levenshtein
|
|
|
|
|
|
| 23 |
transformers
|
| 24 |
torch
|
| 25 |
optimum[onnxruntime]
|
| 26 |
+
Levenshtein
|
| 27 |
+
deepspeed
|
src/AI_Models/wave2vec_inference.py
CHANGED
|
@@ -1,63 +1,416 @@
|
|
| 1 |
-
import torch
|
| 2 |
-
from transformers import (
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
)
|
| 8 |
-
import onnxruntime as rt
|
| 9 |
-
import numpy as np
|
| 10 |
-
import librosa
|
| 11 |
-
import warnings
|
| 12 |
-
import os
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
|
| 14 |
-
warnings.filterwarnings("ignore")
|
| 15 |
|
| 16 |
-
#
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
| 27 |
-
#
|
| 28 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
|
|
|
|
|
|
|
| 30 |
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
return WAVE2VEC2_MODELS.copy()
|
| 34 |
|
|
|
|
|
|
|
| 35 |
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 54 |
|
| 55 |
|
| 56 |
class Wave2Vec2Inference:
|
| 57 |
-
def __init__(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
# Get the actual model name using helper function
|
| 59 |
self.model_name = get_model_name(model_name)
|
| 60 |
-
|
|
|
|
| 61 |
# Auto-detect device
|
| 62 |
if use_gpu:
|
| 63 |
if torch.backends.mps.is_available():
|
|
@@ -71,10 +424,11 @@ class Wave2Vec2Inference:
|
|
| 71 |
|
| 72 |
print(f"Using device: {self.device}")
|
| 73 |
print(f"Loading model: {self.model_name}")
|
|
|
|
| 74 |
|
| 75 |
# Check if model is XLSR and use appropriate processor/model
|
| 76 |
is_xlsr = "xlsr" in self.model_name.lower()
|
| 77 |
-
|
| 78 |
if is_xlsr:
|
| 79 |
print("Using Wav2Vec2Processor and Wav2Vec2ForCTC for XLSR model")
|
| 80 |
self.processor = Wav2Vec2Processor.from_pretrained(self.model_name)
|
|
@@ -83,22 +437,77 @@ class Wave2Vec2Inference:
|
|
| 83 |
print("Using AutoProcessor and AutoModelForCTC")
|
| 84 |
self.processor = AutoProcessor.from_pretrained(self.model_name)
|
| 85 |
self.model = AutoModelForCTC.from_pretrained(self.model_name)
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
self.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 89 |
|
| 90 |
# Disable gradients for inference
|
| 91 |
torch.set_grad_enabled(False)
|
| 92 |
|
| 93 |
-
def
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 94 |
if len(audio_buffer) == 0:
|
| 95 |
return ""
|
| 96 |
|
| 97 |
# Convert to tensor
|
| 98 |
if isinstance(audio_buffer, np.ndarray):
|
| 99 |
audio_tensor = torch.from_numpy(audio_buffer).float()
|
| 100 |
-
|
| 101 |
audio_tensor = torch.tensor(audio_buffer, dtype=torch.float32)
|
|
|
|
|
|
|
| 102 |
|
| 103 |
# Process audio
|
| 104 |
inputs = self.processor(
|
|
@@ -116,12 +525,21 @@ class Wave2Vec2Inference:
|
|
| 116 |
else None
|
| 117 |
)
|
| 118 |
|
|
|
|
|
|
|
|
|
|
| 119 |
# Inference
|
| 120 |
with torch.no_grad():
|
| 121 |
if attention_mask is not None:
|
| 122 |
-
|
| 123 |
else:
|
| 124 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 125 |
|
| 126 |
# Decode
|
| 127 |
predicted_ids = torch.argmax(logits, dim=-1)
|
|
@@ -131,7 +549,16 @@ class Wave2Vec2Inference:
|
|
| 131 |
transcription = self.processor.batch_decode(predicted_ids)[0]
|
| 132 |
return transcription.lower().strip()
|
| 133 |
|
| 134 |
-
def file_to_text(self, filename):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 135 |
try:
|
| 136 |
audio_input, _ = librosa.load(filename, sr=16000, dtype=np.float32)
|
| 137 |
return self.buffer_to_text(audio_input)
|
|
@@ -139,232 +566,101 @@ class Wave2Vec2Inference:
|
|
| 139 |
print(f"Error loading audio file {filename}: {e}")
|
| 140 |
return ""
|
| 141 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 142 |
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
# Always use Wav2Vec2Processor for ONNX (works for all models)
|
| 150 |
-
self.processor = Wav2Vec2Processor.from_pretrained(self.model_name)
|
| 151 |
-
|
| 152 |
-
# Setup ONNX Runtime
|
| 153 |
-
options = rt.SessionOptions()
|
| 154 |
-
options.graph_optimization_level = rt.GraphOptimizationLevel.ORT_ENABLE_ALL
|
| 155 |
-
|
| 156 |
-
# Choose providers based on GPU availability
|
| 157 |
-
providers = []
|
| 158 |
-
if use_gpu and rt.get_available_providers():
|
| 159 |
-
if "CUDAExecutionProvider" in rt.get_available_providers():
|
| 160 |
-
providers.append("CUDAExecutionProvider")
|
| 161 |
-
providers.append("CPUExecutionProvider")
|
| 162 |
|
| 163 |
-
|
| 164 |
-
|
| 165 |
-
print(f"ONNX model loaded with providers: {self.model.get_providers()}")
|
| 166 |
|
| 167 |
-
|
|
|
|
|
|
|
| 168 |
if len(audio_buffer) == 0:
|
| 169 |
-
return ""
|
| 170 |
|
| 171 |
# Convert to tensor
|
| 172 |
if isinstance(audio_buffer, np.ndarray):
|
| 173 |
audio_tensor = torch.from_numpy(audio_buffer).float()
|
| 174 |
else:
|
| 175 |
-
audio_tensor =
|
| 176 |
|
| 177 |
# Process audio
|
| 178 |
inputs = self.processor(
|
| 179 |
audio_tensor,
|
| 180 |
sampling_rate=16_000,
|
| 181 |
-
return_tensors="
|
| 182 |
padding=True,
|
| 183 |
)
|
| 184 |
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 188 |
|
| 189 |
-
#
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
|
|
|
|
|
|
|
| 193 |
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
|
| 198 |
-
except Exception as e:
|
| 199 |
-
print(f"Error loading audio file {filename}: {e}")
|
| 200 |
-
return ""
|
| 201 |
|
|
|
|
|
|
|
|
|
|
| 202 |
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
model = Wav2Vec2ForCTC.from_pretrained(model_id_or_path)
|
| 207 |
-
model.eval()
|
| 208 |
-
|
| 209 |
-
# Create dummy input
|
| 210 |
-
audio_len = 250000
|
| 211 |
-
dummy_input = torch.randn(1, audio_len, requires_grad=True)
|
| 212 |
-
|
| 213 |
-
torch.onnx.export(
|
| 214 |
-
model,
|
| 215 |
-
dummy_input,
|
| 216 |
-
onnx_model_name,
|
| 217 |
-
export_params=True,
|
| 218 |
-
opset_version=14,
|
| 219 |
-
do_constant_folding=True,
|
| 220 |
-
input_names=["input"],
|
| 221 |
-
output_names=["output"],
|
| 222 |
-
dynamic_axes={
|
| 223 |
-
"input": {1: "audio_len"},
|
| 224 |
-
"output": {1: "audio_len"},
|
| 225 |
-
},
|
| 226 |
-
)
|
| 227 |
-
print(f"ONNX model saved to: {onnx_model_name}")
|
| 228 |
-
|
| 229 |
-
|
| 230 |
-
def quantize_onnx_model(onnx_model_path, quantized_model_path):
|
| 231 |
-
"""Quantize ONNX model for faster inference"""
|
| 232 |
-
print("Starting quantization...")
|
| 233 |
-
from onnxruntime.quantization import quantize_dynamic, QuantType
|
| 234 |
-
|
| 235 |
-
quantize_dynamic(
|
| 236 |
-
onnx_model_path, quantized_model_path, weight_type=QuantType.QUInt8
|
| 237 |
-
)
|
| 238 |
-
print(f"Quantized model saved to: {quantized_model_path}")
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
def export_to_onnx(model_name, quantize=False):
|
| 242 |
-
"""
|
| 243 |
-
Export model to ONNX format with optional quantization
|
| 244 |
-
|
| 245 |
-
Args:
|
| 246 |
-
model_name: HuggingFace model name
|
| 247 |
-
quantize: Whether to also create quantized version
|
| 248 |
-
|
| 249 |
-
Returns:
|
| 250 |
-
tuple: (onnx_path, quantized_path or None)
|
| 251 |
-
"""
|
| 252 |
-
onnx_filename = f"{model_name.split('/')[-1]}.onnx"
|
| 253 |
-
convert_to_onnx(model_name, onnx_filename)
|
| 254 |
-
|
| 255 |
-
quantized_path = None
|
| 256 |
-
if quantize:
|
| 257 |
-
quantized_path = onnx_filename.replace(".onnx", ".quantized.onnx")
|
| 258 |
-
quantize_onnx_model(onnx_filename, quantized_path)
|
| 259 |
-
|
| 260 |
-
return onnx_filename, quantized_path
|
| 261 |
-
|
| 262 |
-
|
| 263 |
-
def create_inference(
|
| 264 |
-
model_name=None, use_onnx=False, onnx_path=None, use_gpu=True, use_onnx_quantize=False
|
| 265 |
-
):
|
| 266 |
-
"""
|
| 267 |
-
Create optimized inference instance
|
| 268 |
-
|
| 269 |
-
Args:
|
| 270 |
-
model_name: Model key from WAVE2VEC2_MODELS or full HuggingFace model name (default: uses DEFAULT_MODEL)
|
| 271 |
-
use_onnx: Whether to use ONNX runtime
|
| 272 |
-
onnx_path: Path to ONNX model file
|
| 273 |
-
use_gpu: Whether to use GPU if available
|
| 274 |
-
use_onnx_quantize: Whether to use quantized ONNX model
|
| 275 |
-
|
| 276 |
-
Returns:
|
| 277 |
-
Inference instance
|
| 278 |
-
"""
|
| 279 |
-
# Get the actual model name
|
| 280 |
-
actual_model_name = get_model_name(model_name)
|
| 281 |
-
|
| 282 |
-
if use_onnx:
|
| 283 |
-
if not onnx_path or not os.path.exists(onnx_path):
|
| 284 |
-
# Convert to ONNX if path not provided or doesn't exist
|
| 285 |
-
onnx_filename = f"{actual_model_name.split('/')[-1]}.onnx"
|
| 286 |
-
convert_to_onnx(actual_model_name, onnx_filename)
|
| 287 |
-
onnx_path = onnx_filename
|
| 288 |
-
|
| 289 |
-
if use_onnx_quantize:
|
| 290 |
-
quantized_path = onnx_path.replace(".onnx", ".quantized.onnx")
|
| 291 |
-
if not os.path.exists(quantized_path):
|
| 292 |
-
quantize_onnx_model(onnx_path, quantized_path)
|
| 293 |
-
onnx_path = quantized_path
|
| 294 |
-
|
| 295 |
-
print(f"Using ONNX model: {onnx_path}")
|
| 296 |
-
return Wave2Vec2ONNXInference(model_name, onnx_path, use_gpu)
|
| 297 |
-
else:
|
| 298 |
-
print("Using PyTorch model")
|
| 299 |
-
return Wave2Vec2Inference(model_name, use_gpu)
|
| 300 |
-
|
| 301 |
-
|
| 302 |
-
if __name__ == "__main__":
|
| 303 |
-
import time
|
| 304 |
-
|
| 305 |
-
# Display available models
|
| 306 |
-
print("Available Wave2Vec2 models:")
|
| 307 |
-
for key, model_name in get_available_models().items():
|
| 308 |
-
print(f" {key}: {model_name}")
|
| 309 |
-
print(f"\nDefault model: {DEFAULT_MODEL}")
|
| 310 |
-
print()
|
| 311 |
-
|
| 312 |
-
# Test with different models
|
| 313 |
-
test_models = ["english_large", "multilingual", "english_960h"]
|
| 314 |
-
test_file = "test.wav"
|
| 315 |
-
|
| 316 |
-
if not os.path.exists(test_file):
|
| 317 |
-
print(f"Test file {test_file} not found. Please provide a valid audio file.")
|
| 318 |
-
print("Creating example usage without actual file...")
|
| 319 |
-
|
| 320 |
-
# Example usage without file
|
| 321 |
-
print("\n=== Example Usage ===")
|
| 322 |
-
|
| 323 |
-
# Using default model
|
| 324 |
-
print("1. Using default model:")
|
| 325 |
-
asr_default = create_inference()
|
| 326 |
-
print(f" Model loaded: {asr_default.model_name}")
|
| 327 |
-
|
| 328 |
-
# Using model key
|
| 329 |
-
print("\n2. Using model key 'english_large':")
|
| 330 |
-
asr_key = create_inference("english_large")
|
| 331 |
-
print(f" Model loaded: {asr_key.model_name}")
|
| 332 |
-
|
| 333 |
-
# Using full model name
|
| 334 |
-
print("\n3. Using full model name:")
|
| 335 |
-
asr_full = create_inference("facebook/wav2vec2-base-960h")
|
| 336 |
-
print(f" Model loaded: {asr_full.model_name}")
|
| 337 |
-
|
| 338 |
-
exit(0)
|
| 339 |
|
| 340 |
-
|
| 341 |
-
|
| 342 |
-
|
| 343 |
-
|
| 344 |
-
|
| 345 |
-
|
| 346 |
-
|
| 347 |
-
|
| 348 |
-
|
| 349 |
-
|
| 350 |
-
|
| 351 |
-
|
| 352 |
-
|
| 353 |
-
|
| 354 |
-
|
| 355 |
-
|
| 356 |
-
|
| 357 |
-
|
| 358 |
-
|
| 359 |
-
# Test performance
|
| 360 |
-
times = []
|
| 361 |
-
for i in range(3):
|
| 362 |
-
start_time = time.time()
|
| 363 |
-
text = asr.file_to_text(test_file)
|
| 364 |
-
end_time = time.time()
|
| 365 |
-
execution_time = end_time - start_time
|
| 366 |
-
times.append(execution_time)
|
| 367 |
-
print(f"Run {i+1}: {execution_time:.3f}s - {text[:50]}...")
|
| 368 |
-
|
| 369 |
-
avg_time = sum(times) / len(times)
|
| 370 |
-
print(f"Average time: {avg_time:.3f}s")
|
|
|
|
| 1 |
+
# import torch
|
| 2 |
+
# from transformers import (
|
| 3 |
+
# AutoModelForCTC,
|
| 4 |
+
# AutoProcessor,
|
| 5 |
+
# Wav2Vec2Processor,
|
| 6 |
+
# Wav2Vec2ForCTC,
|
| 7 |
+
# )
|
| 8 |
+
# import onnxruntime as rt
|
| 9 |
+
# import numpy as np
|
| 10 |
+
# import librosa
|
| 11 |
+
# import warnings
|
| 12 |
+
# import os
|
| 13 |
+
|
| 14 |
+
# warnings.filterwarnings("ignore")
|
| 15 |
+
|
| 16 |
+
# # Available Wave2Vec2 models
|
| 17 |
+
# WAVE2VEC2_MODELS = {
|
| 18 |
+
# "english_large": "jonatasgrosman/wav2vec2-large-xlsr-53-english",
|
| 19 |
+
# "multilingual": "facebook/wav2vec2-large-xlsr-53",
|
| 20 |
+
# "english_960h": "facebook/wav2vec2-large-960h-lv60-self",
|
| 21 |
+
# "base_english": "facebook/wav2vec2-base-960h",
|
| 22 |
+
# "large_english": "facebook/wav2vec2-large-960h",
|
| 23 |
+
# "xlsr_english": "jonatasgrosman/wav2vec2-large-xlsr-53-english",
|
| 24 |
+
# "xlsr_multilingual": "facebook/wav2vec2-large-xlsr-53"
|
| 25 |
+
# }
|
| 26 |
+
|
| 27 |
+
# # Default model
|
| 28 |
+
# DEFAULT_MODEL = "jonatasgrosman/wav2vec2-large-xlsr-53-english"
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
# def get_available_models():
|
| 32 |
+
# """Return dictionary of available Wave2Vec2 models"""
|
| 33 |
+
# return WAVE2VEC2_MODELS.copy()
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
# def get_model_name(model_key=None):
|
| 37 |
+
# """
|
| 38 |
+
# Get model name from key or return default
|
| 39 |
+
|
| 40 |
+
# Args:
|
| 41 |
+
# model_key: Key from WAVE2VEC2_MODELS or full model name
|
| 42 |
+
|
| 43 |
+
# Returns:
|
| 44 |
+
# str: Full model name
|
| 45 |
+
# """
|
| 46 |
+
# if model_key is None:
|
| 47 |
+
# return DEFAULT_MODEL
|
| 48 |
+
|
| 49 |
+
# if model_key in WAVE2VEC2_MODELS:
|
| 50 |
+
# return WAVE2VEC2_MODELS[model_key]
|
| 51 |
+
|
| 52 |
+
# # If it's already a full model name, return as is
|
| 53 |
+
# return model_key
|
| 54 |
|
|
|
|
| 55 |
|
| 56 |
+
# class Wave2Vec2Inference:
|
| 57 |
+
# def __init__(self, model_name=None, use_gpu=True):
|
| 58 |
+
# # Get the actual model name using helper function
|
| 59 |
+
# self.model_name = get_model_name(model_name)
|
| 60 |
+
|
| 61 |
+
# # Auto-detect device
|
| 62 |
+
# if use_gpu:
|
| 63 |
+
# if torch.backends.mps.is_available():
|
| 64 |
+
# self.device = "mps"
|
| 65 |
+
# elif torch.cuda.is_available():
|
| 66 |
+
# self.device = "cuda"
|
| 67 |
+
# else:
|
| 68 |
+
# self.device = "cpu"
|
| 69 |
+
# else:
|
| 70 |
+
# self.device = "cpu"
|
| 71 |
+
|
| 72 |
+
# print(f"Using device: {self.device}")
|
| 73 |
+
# print(f"Loading model: {self.model_name}")
|
| 74 |
+
|
| 75 |
+
# # Check if model is XLSR and use appropriate processor/model
|
| 76 |
+
# is_xlsr = "xlsr" in self.model_name.lower()
|
| 77 |
+
|
| 78 |
+
# if is_xlsr:
|
| 79 |
+
# print("Using Wav2Vec2Processor and Wav2Vec2ForCTC for XLSR model")
|
| 80 |
+
# self.processor = Wav2Vec2Processor.from_pretrained(self.model_name)
|
| 81 |
+
# self.model = Wav2Vec2ForCTC.from_pretrained(self.model_name)
|
| 82 |
+
# else:
|
| 83 |
+
# print("Using AutoProcessor and AutoModelForCTC")
|
| 84 |
+
# self.processor = AutoProcessor.from_pretrained(self.model_name)
|
| 85 |
+
# self.model = AutoModelForCTC.from_pretrained(self.model_name)
|
| 86 |
+
|
| 87 |
+
# self.model.to(self.device)
|
| 88 |
+
# self.model.eval()
|
| 89 |
+
|
| 90 |
+
# # Disable gradients for inference
|
| 91 |
+
# torch.set_grad_enabled(False)
|
| 92 |
+
|
| 93 |
+
# def buffer_to_text(self, audio_buffer):
|
| 94 |
+
# if len(audio_buffer) == 0:
|
| 95 |
+
# return ""
|
| 96 |
+
|
| 97 |
+
# # Convert to tensor
|
| 98 |
+
# if isinstance(audio_buffer, np.ndarray):
|
| 99 |
+
# audio_tensor = torch.from_numpy(audio_buffer).float()
|
| 100 |
+
# else:
|
| 101 |
+
# audio_tensor = torch.tensor(audio_buffer, dtype=torch.float32)
|
| 102 |
+
|
| 103 |
+
# # Process audio
|
| 104 |
+
# inputs = self.processor(
|
| 105 |
+
# audio_tensor,
|
| 106 |
+
# sampling_rate=16_000,
|
| 107 |
+
# return_tensors="pt",
|
| 108 |
+
# padding=True,
|
| 109 |
+
# )
|
| 110 |
+
|
| 111 |
+
# # Move to device
|
| 112 |
+
# input_values = inputs.input_values.to(self.device)
|
| 113 |
+
# attention_mask = (
|
| 114 |
+
# inputs.attention_mask.to(self.device)
|
| 115 |
+
# if "attention_mask" in inputs
|
| 116 |
+
# else None
|
| 117 |
+
# )
|
| 118 |
+
|
| 119 |
+
# # Inference
|
| 120 |
+
# with torch.no_grad():
|
| 121 |
+
# if attention_mask is not None:
|
| 122 |
+
# logits = self.model(input_values, attention_mask=attention_mask).logits
|
| 123 |
+
# else:
|
| 124 |
+
# logits = self.model(input_values).logits
|
| 125 |
+
|
| 126 |
+
# # Decode
|
| 127 |
+
# predicted_ids = torch.argmax(logits, dim=-1)
|
| 128 |
+
# if self.device != "cpu":
|
| 129 |
+
# predicted_ids = predicted_ids.cpu()
|
| 130 |
+
|
| 131 |
+
# transcription = self.processor.batch_decode(predicted_ids)[0]
|
| 132 |
+
# return transcription.lower().strip()
|
| 133 |
+
|
| 134 |
+
# def file_to_text(self, filename):
|
| 135 |
+
# try:
|
| 136 |
+
# audio_input, _ = librosa.load(filename, sr=16000, dtype=np.float32)
|
| 137 |
+
# return self.buffer_to_text(audio_input)
|
| 138 |
+
# except Exception as e:
|
| 139 |
+
# print(f"Error loading audio file {filename}: {e}")
|
| 140 |
+
# return ""
|
| 141 |
+
|
| 142 |
+
|
| 143 |
+
# class Wave2Vec2ONNXInference:
|
| 144 |
+
# def __init__(self, model_name=None, onnx_path=None, use_gpu=True):
|
| 145 |
+
# # Get the actual model name using helper function
|
| 146 |
+
# self.model_name = get_model_name(model_name)
|
| 147 |
+
# print(f"Loading ONNX model: {self.model_name}")
|
| 148 |
+
|
| 149 |
+
# # Always use Wav2Vec2Processor for ONNX (works for all models)
|
| 150 |
+
# self.processor = Wav2Vec2Processor.from_pretrained(self.model_name)
|
| 151 |
+
|
| 152 |
+
# # Setup ONNX Runtime
|
| 153 |
+
# options = rt.SessionOptions()
|
| 154 |
+
# options.graph_optimization_level = rt.GraphOptimizationLevel.ORT_ENABLE_ALL
|
| 155 |
+
|
| 156 |
+
# # Choose providers based on GPU availability
|
| 157 |
+
# providers = []
|
| 158 |
+
# if use_gpu and rt.get_available_providers():
|
| 159 |
+
# if "CUDAExecutionProvider" in rt.get_available_providers():
|
| 160 |
+
# providers.append("CUDAExecutionProvider")
|
| 161 |
+
# providers.append("CPUExecutionProvider")
|
| 162 |
+
|
| 163 |
+
# self.model = rt.InferenceSession(onnx_path, options, providers=providers)
|
| 164 |
+
# self.input_name = self.model.get_inputs()[0].name
|
| 165 |
+
# print(f"ONNX model loaded with providers: {self.model.get_providers()}")
|
| 166 |
+
|
| 167 |
+
# def buffer_to_text(self, audio_buffer):
|
| 168 |
+
# if len(audio_buffer) == 0:
|
| 169 |
+
# return ""
|
| 170 |
+
|
| 171 |
+
# # Convert to tensor
|
| 172 |
+
# if isinstance(audio_buffer, np.ndarray):
|
| 173 |
+
# audio_tensor = torch.from_numpy(audio_buffer).float()
|
| 174 |
+
# else:
|
| 175 |
+
# audio_tensor = torch.tensor(audio_buffer, dtype=torch.float32)
|
| 176 |
+
|
| 177 |
+
# # Process audio
|
| 178 |
+
# inputs = self.processor(
|
| 179 |
+
# audio_tensor,
|
| 180 |
+
# sampling_rate=16_000,
|
| 181 |
+
# return_tensors="np",
|
| 182 |
+
# padding=True,
|
| 183 |
+
# )
|
| 184 |
+
|
| 185 |
+
# # ONNX inference
|
| 186 |
+
# input_values = inputs.input_values.astype(np.float32)
|
| 187 |
+
# onnx_outputs = self.model.run(None, {self.input_name: input_values})[0]
|
| 188 |
+
|
| 189 |
+
# # Decode
|
| 190 |
+
# prediction = np.argmax(onnx_outputs, axis=-1)
|
| 191 |
+
# transcription = self.processor.decode(prediction.squeeze().tolist())
|
| 192 |
+
# return transcription.lower().strip()
|
| 193 |
+
|
| 194 |
+
# def file_to_text(self, filename):
|
| 195 |
+
# try:
|
| 196 |
+
# audio_input, _ = librosa.load(filename, sr=16000, dtype=np.float32)
|
| 197 |
+
# return self.buffer_to_text(audio_input)
|
| 198 |
+
# except Exception as e:
|
| 199 |
+
# print(f"Error loading audio file {filename}: {e}")
|
| 200 |
+
# return ""
|
| 201 |
+
|
| 202 |
+
|
| 203 |
+
# def convert_to_onnx(model_id_or_path, onnx_model_name):
|
| 204 |
+
# """Convert PyTorch model to ONNX format"""
|
| 205 |
+
# print(f"Converting {model_id_or_path} to ONNX...")
|
| 206 |
+
# model = Wav2Vec2ForCTC.from_pretrained(model_id_or_path)
|
| 207 |
+
# model.eval()
|
| 208 |
+
|
| 209 |
+
# # Create dummy input
|
| 210 |
+
# audio_len = 250000
|
| 211 |
+
# dummy_input = torch.randn(1, audio_len, requires_grad=True)
|
| 212 |
+
|
| 213 |
+
# torch.onnx.export(
|
| 214 |
+
# model,
|
| 215 |
+
# dummy_input,
|
| 216 |
+
# onnx_model_name,
|
| 217 |
+
# export_params=True,
|
| 218 |
+
# opset_version=14,
|
| 219 |
+
# do_constant_folding=True,
|
| 220 |
+
# input_names=["input"],
|
| 221 |
+
# output_names=["output"],
|
| 222 |
+
# dynamic_axes={
|
| 223 |
+
# "input": {1: "audio_len"},
|
| 224 |
+
# "output": {1: "audio_len"},
|
| 225 |
+
# },
|
| 226 |
+
# )
|
| 227 |
+
# print(f"ONNX model saved to: {onnx_model_name}")
|
| 228 |
+
|
| 229 |
+
|
| 230 |
+
# def quantize_onnx_model(onnx_model_path, quantized_model_path):
|
| 231 |
+
# """Quantize ONNX model for faster inference"""
|
| 232 |
+
# print("Starting quantization...")
|
| 233 |
+
# from onnxruntime.quantization import quantize_dynamic, QuantType
|
| 234 |
+
|
| 235 |
+
# quantize_dynamic(
|
| 236 |
+
# onnx_model_path, quantized_model_path, weight_type=QuantType.QUInt8
|
| 237 |
+
# )
|
| 238 |
+
# print(f"Quantized model saved to: {quantized_model_path}")
|
| 239 |
+
|
| 240 |
+
|
| 241 |
+
# def export_to_onnx(model_name, quantize=False):
|
| 242 |
+
# """
|
| 243 |
+
# Export model to ONNX format with optional quantization
|
| 244 |
+
|
| 245 |
+
# Args:
|
| 246 |
+
# model_name: HuggingFace model name
|
| 247 |
+
# quantize: Whether to also create quantized version
|
| 248 |
+
|
| 249 |
+
# Returns:
|
| 250 |
+
# tuple: (onnx_path, quantized_path or None)
|
| 251 |
+
# """
|
| 252 |
+
# onnx_filename = f"{model_name.split('/')[-1]}.onnx"
|
| 253 |
+
# convert_to_onnx(model_name, onnx_filename)
|
| 254 |
+
|
| 255 |
+
# quantized_path = None
|
| 256 |
+
# if quantize:
|
| 257 |
+
# quantized_path = onnx_filename.replace(".onnx", ".quantized.onnx")
|
| 258 |
+
# quantize_onnx_model(onnx_filename, quantized_path)
|
| 259 |
+
|
| 260 |
+
# return onnx_filename, quantized_path
|
| 261 |
+
|
| 262 |
+
|
| 263 |
+
# def create_inference(
|
| 264 |
+
# model_name=None, use_onnx=False, onnx_path=None, use_gpu=True, use_onnx_quantize=False
|
| 265 |
+
# ):
|
| 266 |
+
# """
|
| 267 |
+
# Create optimized inference instance
|
| 268 |
+
|
| 269 |
+
# Args:
|
| 270 |
+
# model_name: Model key from WAVE2VEC2_MODELS or full HuggingFace model name (default: uses DEFAULT_MODEL)
|
| 271 |
+
# use_onnx: Whether to use ONNX runtime
|
| 272 |
+
# onnx_path: Path to ONNX model file
|
| 273 |
+
# use_gpu: Whether to use GPU if available
|
| 274 |
+
# use_onnx_quantize: Whether to use quantized ONNX model
|
| 275 |
+
|
| 276 |
+
# Returns:
|
| 277 |
+
# Inference instance
|
| 278 |
+
# """
|
| 279 |
+
# # Get the actual model name
|
| 280 |
+
# actual_model_name = get_model_name(model_name)
|
| 281 |
+
|
| 282 |
+
# if use_onnx:
|
| 283 |
+
# if not onnx_path or not os.path.exists(onnx_path):
|
| 284 |
+
# # Convert to ONNX if path not provided or doesn't exist
|
| 285 |
+
# onnx_filename = f"{actual_model_name.split('/')[-1]}.onnx"
|
| 286 |
+
# convert_to_onnx(actual_model_name, onnx_filename)
|
| 287 |
+
# onnx_path = onnx_filename
|
| 288 |
+
|
| 289 |
+
# if use_onnx_quantize:
|
| 290 |
+
# quantized_path = onnx_path.replace(".onnx", ".quantized.onnx")
|
| 291 |
+
# if not os.path.exists(quantized_path):
|
| 292 |
+
# quantize_onnx_model(onnx_path, quantized_path)
|
| 293 |
+
# onnx_path = quantized_path
|
| 294 |
+
|
| 295 |
+
# print(f"Using ONNX model: {onnx_path}")
|
| 296 |
+
# return Wave2Vec2ONNXInference(model_name, onnx_path, use_gpu)
|
| 297 |
+
# else:
|
| 298 |
+
# print("Using PyTorch model")
|
| 299 |
+
# return Wave2Vec2Inference(model_name, use_gpu)
|
| 300 |
+
|
| 301 |
+
|
| 302 |
+
# if __name__ == "__main__":
|
| 303 |
+
# import time
|
| 304 |
+
|
| 305 |
+
# # Display available models
|
| 306 |
+
# print("Available Wave2Vec2 models:")
|
| 307 |
+
# for key, model_name in get_available_models().items():
|
| 308 |
+
# print(f" {key}: {model_name}")
|
| 309 |
+
# print(f"\nDefault model: {DEFAULT_MODEL}")
|
| 310 |
+
# print()
|
| 311 |
+
|
| 312 |
+
# # Test with different models
|
| 313 |
+
# test_models = ["english_large", "multilingual", "english_960h"]
|
| 314 |
+
# test_file = "test.wav"
|
| 315 |
+
|
| 316 |
+
# if not os.path.exists(test_file):
|
| 317 |
+
# print(f"Test file {test_file} not found. Please provide a valid audio file.")
|
| 318 |
+
# print("Creating example usage without actual file...")
|
| 319 |
+
|
| 320 |
+
# # Example usage without file
|
| 321 |
+
# print("\n=== Example Usage ===")
|
| 322 |
+
|
| 323 |
+
# # Using default model
|
| 324 |
+
# print("1. Using default model:")
|
| 325 |
+
# asr_default = create_inference()
|
| 326 |
+
# print(f" Model loaded: {asr_default.model_name}")
|
| 327 |
+
|
| 328 |
+
# # Using model key
|
| 329 |
+
# print("\n2. Using model key 'english_large':")
|
| 330 |
+
# asr_key = create_inference("english_large")
|
| 331 |
+
# print(f" Model loaded: {asr_key.model_name}")
|
| 332 |
+
|
| 333 |
+
# # Using full model name
|
| 334 |
+
# print("\n3. Using full model name:")
|
| 335 |
+
# asr_full = create_inference("facebook/wav2vec2-base-960h")
|
| 336 |
+
# print(f" Model loaded: {asr_full.model_name}")
|
| 337 |
+
|
| 338 |
+
# exit(0)
|
| 339 |
|
| 340 |
+
# # Test different model configurations
|
| 341 |
+
# for model_key in test_models:
|
| 342 |
+
# print(f"\n=== Testing model: {model_key} ===")
|
| 343 |
+
|
| 344 |
+
# # Test different configurations
|
| 345 |
+
# configs = [
|
| 346 |
+
# {"use_onnx": False, "use_gpu": True},
|
| 347 |
+
# {"use_onnx": True, "use_gpu": True, "use_onnx_quantize": False},
|
| 348 |
+
# ]
|
| 349 |
|
| 350 |
+
# for config in configs:
|
| 351 |
+
# print(f"\nConfig: {config}")
|
| 352 |
|
| 353 |
+
# # Create inference instance with model selection
|
| 354 |
+
# asr = create_inference(model_key, **config)
|
|
|
|
| 355 |
|
| 356 |
+
# # Warm up
|
| 357 |
+
# asr.file_to_text(test_file)
|
| 358 |
|
| 359 |
+
# # Test performance
|
| 360 |
+
# times = []
|
| 361 |
+
# for i in range(3):
|
| 362 |
+
# start_time = time.time()
|
| 363 |
+
# text = asr.file_to_text(test_file)
|
| 364 |
+
# end_time = time.time()
|
| 365 |
+
# execution_time = end_time - start_time
|
| 366 |
+
# times.append(execution_time)
|
| 367 |
+
# print(f"Run {i+1}: {execution_time:.3f}s - {text[:50]}...")
|
| 368 |
+
|
| 369 |
+
# avg_time = sum(times) / len(times)
|
| 370 |
+
# print(f"Average time: {avg_time:.3f}s")
|
| 371 |
+
|
| 372 |
+
|
| 373 |
+
|
| 374 |
+
import torch
|
| 375 |
+
from transformers import (
|
| 376 |
+
Wav2Vec2ForCTC,
|
| 377 |
+
Wav2Vec2Processor,
|
| 378 |
+
AutoProcessor,
|
| 379 |
+
AutoModelForCTC,
|
| 380 |
+
)
|
| 381 |
+
|
| 382 |
+
import deepspeed
|
| 383 |
+
import librosa
|
| 384 |
+
import numpy as np
|
| 385 |
+
from typing import Optional, List, Union
|
| 386 |
+
|
| 387 |
+
|
| 388 |
+
def get_model_name(model_name: Optional[str] = None) -> str:
|
| 389 |
+
"""Helper function to get model name with default fallback"""
|
| 390 |
+
if model_name is None:
|
| 391 |
+
return "facebook/wav2vec2-large-robust-ft-libri-960h"
|
| 392 |
+
return model_name
|
| 393 |
|
| 394 |
|
| 395 |
class Wave2Vec2Inference:
|
| 396 |
+
def __init__(
|
| 397 |
+
self,
|
| 398 |
+
model_name: Optional[str] = None,
|
| 399 |
+
use_gpu: bool = True,
|
| 400 |
+
use_deepspeed: bool = True,
|
| 401 |
+
):
|
| 402 |
+
"""
|
| 403 |
+
Initialize Wav2Vec2 model for inference with optional DeepSpeed optimization.
|
| 404 |
+
|
| 405 |
+
Args:
|
| 406 |
+
model_name: HuggingFace model name or None for default
|
| 407 |
+
use_gpu: Whether to use GPU acceleration
|
| 408 |
+
use_deepspeed: Whether to use DeepSpeed optimization
|
| 409 |
+
"""
|
| 410 |
# Get the actual model name using helper function
|
| 411 |
self.model_name = get_model_name(model_name)
|
| 412 |
+
self.use_deepspeed = use_deepspeed
|
| 413 |
+
|
| 414 |
# Auto-detect device
|
| 415 |
if use_gpu:
|
| 416 |
if torch.backends.mps.is_available():
|
|
|
|
| 424 |
|
| 425 |
print(f"Using device: {self.device}")
|
| 426 |
print(f"Loading model: {self.model_name}")
|
| 427 |
+
print(f"DeepSpeed enabled: {self.use_deepspeed}")
|
| 428 |
|
| 429 |
# Check if model is XLSR and use appropriate processor/model
|
| 430 |
is_xlsr = "xlsr" in self.model_name.lower()
|
| 431 |
+
|
| 432 |
if is_xlsr:
|
| 433 |
print("Using Wav2Vec2Processor and Wav2Vec2ForCTC for XLSR model")
|
| 434 |
self.processor = Wav2Vec2Processor.from_pretrained(self.model_name)
|
|
|
|
| 437 |
print("Using AutoProcessor and AutoModelForCTC")
|
| 438 |
self.processor = AutoProcessor.from_pretrained(self.model_name)
|
| 439 |
self.model = AutoModelForCTC.from_pretrained(self.model_name)
|
| 440 |
+
|
| 441 |
+
# Initialize DeepSpeed if enabled
|
| 442 |
+
if self.use_deepspeed:
|
| 443 |
+
self._init_deepspeed()
|
| 444 |
+
else:
|
| 445 |
+
self.model.to(self.device)
|
| 446 |
+
self.model.eval()
|
| 447 |
+
self.ds_engine = None
|
| 448 |
|
| 449 |
# Disable gradients for inference
|
| 450 |
torch.set_grad_enabled(False)
|
| 451 |
|
| 452 |
+
def _init_deepspeed(self):
|
| 453 |
+
"""Initialize DeepSpeed inference engine"""
|
| 454 |
+
try:
|
| 455 |
+
# DeepSpeed configuration based on device
|
| 456 |
+
if self.device == "cuda":
|
| 457 |
+
ds_config = {
|
| 458 |
+
"tensor_parallel": {"tp_size": 1},
|
| 459 |
+
"dtype": torch.float32,
|
| 460 |
+
"replace_with_kernel_inject": True,
|
| 461 |
+
"enable_cuda_graph": False,
|
| 462 |
+
}
|
| 463 |
+
else:
|
| 464 |
+
ds_config = {
|
| 465 |
+
"tensor_parallel": {"tp_size": 1},
|
| 466 |
+
"dtype": torch.float32,
|
| 467 |
+
"replace_with_kernel_inject": False,
|
| 468 |
+
"enable_cuda_graph": False,
|
| 469 |
+
}
|
| 470 |
+
|
| 471 |
+
print("Initializing DeepSpeed inference engine...")
|
| 472 |
+
self.ds_engine = deepspeed.init_inference(self.model, **ds_config)
|
| 473 |
+
self.ds_engine.module.to(self.device)
|
| 474 |
+
|
| 475 |
+
except Exception as e:
|
| 476 |
+
print(f"DeepSpeed initialization failed: {e}")
|
| 477 |
+
print("Falling back to standard PyTorch inference...")
|
| 478 |
+
self.use_deepspeed = False
|
| 479 |
+
self.ds_engine = None
|
| 480 |
+
self.model.to(self.device)
|
| 481 |
+
self.model.eval()
|
| 482 |
+
|
| 483 |
+
def _get_model(self):
|
| 484 |
+
"""Get the appropriate model for inference"""
|
| 485 |
+
if self.use_deepspeed and self.ds_engine is not None:
|
| 486 |
+
return self.ds_engine.module
|
| 487 |
+
return self.model
|
| 488 |
+
|
| 489 |
+
def buffer_to_text(
|
| 490 |
+
self, audio_buffer: Union[np.ndarray, torch.Tensor, List]
|
| 491 |
+
) -> str:
|
| 492 |
+
"""
|
| 493 |
+
Convert audio buffer to text transcription.
|
| 494 |
+
|
| 495 |
+
Args:
|
| 496 |
+
audio_buffer: Audio data as numpy array, tensor, or list
|
| 497 |
+
|
| 498 |
+
Returns:
|
| 499 |
+
str: Transcribed text
|
| 500 |
+
"""
|
| 501 |
if len(audio_buffer) == 0:
|
| 502 |
return ""
|
| 503 |
|
| 504 |
# Convert to tensor
|
| 505 |
if isinstance(audio_buffer, np.ndarray):
|
| 506 |
audio_tensor = torch.from_numpy(audio_buffer).float()
|
| 507 |
+
elif isinstance(audio_buffer, list):
|
| 508 |
audio_tensor = torch.tensor(audio_buffer, dtype=torch.float32)
|
| 509 |
+
else:
|
| 510 |
+
audio_tensor = audio_buffer.float()
|
| 511 |
|
| 512 |
# Process audio
|
| 513 |
inputs = self.processor(
|
|
|
|
| 525 |
else None
|
| 526 |
)
|
| 527 |
|
| 528 |
+
# Get the appropriate model
|
| 529 |
+
model = self._get_model()
|
| 530 |
+
|
| 531 |
# Inference
|
| 532 |
with torch.no_grad():
|
| 533 |
if attention_mask is not None:
|
| 534 |
+
outputs = model(input_values, attention_mask=attention_mask)
|
| 535 |
else:
|
| 536 |
+
outputs = model(input_values)
|
| 537 |
+
|
| 538 |
+
# Handle different output formats
|
| 539 |
+
if hasattr(outputs, "logits"):
|
| 540 |
+
logits = outputs.logits
|
| 541 |
+
else:
|
| 542 |
+
logits = outputs
|
| 543 |
|
| 544 |
# Decode
|
| 545 |
predicted_ids = torch.argmax(logits, dim=-1)
|
|
|
|
| 549 |
transcription = self.processor.batch_decode(predicted_ids)[0]
|
| 550 |
return transcription.lower().strip()
|
| 551 |
|
| 552 |
+
def file_to_text(self, filename: str) -> str:
|
| 553 |
+
"""
|
| 554 |
+
Transcribe audio file to text.
|
| 555 |
+
|
| 556 |
+
Args:
|
| 557 |
+
filename: Path to audio file
|
| 558 |
+
|
| 559 |
+
Returns:
|
| 560 |
+
str: Transcribed text
|
| 561 |
+
"""
|
| 562 |
try:
|
| 563 |
audio_input, _ = librosa.load(filename, sr=16000, dtype=np.float32)
|
| 564 |
return self.buffer_to_text(audio_input)
|
|
|
|
| 566 |
print(f"Error loading audio file {filename}: {e}")
|
| 567 |
return ""
|
| 568 |
|
| 569 |
+
def batch_file_to_text(self, filenames: List[str]) -> List[str]:
|
| 570 |
+
"""
|
| 571 |
+
Transcribe multiple audio files to text.
|
| 572 |
+
|
| 573 |
+
Args:
|
| 574 |
+
filenames: List of audio file paths
|
| 575 |
+
|
| 576 |
+
Returns:
|
| 577 |
+
List[str]: List of transcribed texts
|
| 578 |
+
"""
|
| 579 |
+
results = []
|
| 580 |
+
for i, filename in enumerate(filenames):
|
| 581 |
+
print(f"Processing file {i+1}/{len(filenames)}: {filename}")
|
| 582 |
+
transcription = self.file_to_text(filename)
|
| 583 |
+
results.append(transcription)
|
| 584 |
+
if transcription:
|
| 585 |
+
print(f"Transcription: {transcription}")
|
| 586 |
+
else:
|
| 587 |
+
print("Failed to transcribe")
|
| 588 |
+
return results
|
| 589 |
|
| 590 |
+
def transcribe_with_confidence(
|
| 591 |
+
self, audio_buffer: Union[np.ndarray, torch.Tensor]
|
| 592 |
+
) -> tuple:
|
| 593 |
+
"""
|
| 594 |
+
Transcribe audio and return confidence scores.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 595 |
|
| 596 |
+
Args:
|
| 597 |
+
audio_buffer: Audio data
|
|
|
|
| 598 |
|
| 599 |
+
Returns:
|
| 600 |
+
tuple: (transcription, confidence_scores)
|
| 601 |
+
"""
|
| 602 |
if len(audio_buffer) == 0:
|
| 603 |
+
return "", []
|
| 604 |
|
| 605 |
# Convert to tensor
|
| 606 |
if isinstance(audio_buffer, np.ndarray):
|
| 607 |
audio_tensor = torch.from_numpy(audio_buffer).float()
|
| 608 |
else:
|
| 609 |
+
audio_tensor = audio_buffer.float()
|
| 610 |
|
| 611 |
# Process audio
|
| 612 |
inputs = self.processor(
|
| 613 |
audio_tensor,
|
| 614 |
sampling_rate=16_000,
|
| 615 |
+
return_tensors="pt",
|
| 616 |
padding=True,
|
| 617 |
)
|
| 618 |
|
| 619 |
+
input_values = inputs.input_values.to(self.device)
|
| 620 |
+
attention_mask = (
|
| 621 |
+
inputs.attention_mask.to(self.device)
|
| 622 |
+
if "attention_mask" in inputs
|
| 623 |
+
else None
|
| 624 |
+
)
|
| 625 |
+
|
| 626 |
+
model = self._get_model()
|
| 627 |
|
| 628 |
+
# Inference
|
| 629 |
+
with torch.no_grad():
|
| 630 |
+
if attention_mask is not None:
|
| 631 |
+
outputs = model(input_values, attention_mask=attention_mask)
|
| 632 |
+
else:
|
| 633 |
+
outputs = model(input_values)
|
| 634 |
|
| 635 |
+
if hasattr(outputs, "logits"):
|
| 636 |
+
logits = outputs.logits
|
| 637 |
+
else:
|
| 638 |
+
logits = outputs
|
|
|
|
|
|
|
|
|
|
| 639 |
|
| 640 |
+
# Get probabilities and confidence scores
|
| 641 |
+
probs = torch.nn.functional.softmax(logits, dim=-1)
|
| 642 |
+
predicted_ids = torch.argmax(logits, dim=-1)
|
| 643 |
|
| 644 |
+
# Calculate confidence as max probability for each prediction
|
| 645 |
+
max_probs = torch.max(probs, dim=-1)[0]
|
| 646 |
+
confidence_scores = max_probs.cpu().numpy().tolist()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 647 |
|
| 648 |
+
if self.device != "cpu":
|
| 649 |
+
predicted_ids = predicted_ids.cpu()
|
| 650 |
+
|
| 651 |
+
transcription = self.processor.batch_decode(predicted_ids)[0]
|
| 652 |
+
return transcription.lower().strip(), confidence_scores
|
| 653 |
+
|
| 654 |
+
def cleanup(self):
|
| 655 |
+
"""Clean up resources"""
|
| 656 |
+
if hasattr(self, "ds_engine") and self.ds_engine is not None:
|
| 657 |
+
del self.ds_engine
|
| 658 |
+
if hasattr(self, "model"):
|
| 659 |
+
del self.model
|
| 660 |
+
if hasattr(self, "processor"):
|
| 661 |
+
del self.processor
|
| 662 |
+
torch.cuda.empty_cache() if torch.cuda.is_available() else None
|
| 663 |
+
|
| 664 |
+
def __del__(self):
|
| 665 |
+
"""Destructor to clean up resources"""
|
| 666 |
+
self.cleanup()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
src/apis/controllers/speaking_controller.py
CHANGED
|
@@ -14,10 +14,12 @@ import Levenshtein
|
|
| 14 |
from dataclasses import dataclass
|
| 15 |
from enum import Enum
|
| 16 |
import os
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
|
|
|
|
|
|
| 21 |
from src.utils.vietnamese_tips import vietnamese_tips
|
| 22 |
|
| 23 |
# Download required NLTK data
|
|
@@ -78,9 +80,7 @@ class EnhancedWav2Vec2CharacterASR:
|
|
| 78 |
export_to_onnx(model_name, quantize=quantized)
|
| 79 |
|
| 80 |
# Use optimized inference
|
| 81 |
-
self.model =
|
| 82 |
-
model_name=model_name, use_onnx=onnx, use_onnx_quantize=quantized
|
| 83 |
-
)
|
| 84 |
|
| 85 |
def transcribe_with_features(self, audio_path: str, retry_count: int = 0) -> Dict:
|
| 86 |
"""Enhanced transcription with audio features for prosody analysis - Optimized with retry mechanism"""
|
|
|
|
| 14 |
from dataclasses import dataclass
|
| 15 |
from enum import Enum
|
| 16 |
import os
|
| 17 |
+
|
| 18 |
+
# from src.AI_Models.wave2vec_inference import (
|
| 19 |
+
# create_inference,
|
| 20 |
+
# export_to_onnx,
|
| 21 |
+
# )
|
| 22 |
+
from src.AI_Models.wave2vec_inference import Wave2Vec2Inference
|
| 23 |
from src.utils.vietnamese_tips import vietnamese_tips
|
| 24 |
|
| 25 |
# Download required NLTK data
|
|
|
|
| 80 |
export_to_onnx(model_name, quantize=quantized)
|
| 81 |
|
| 82 |
# Use optimized inference
|
| 83 |
+
self.model = Wave2Vec2Inference(model_name, use_gpu=False, use_deepspeed=True)
|
|
|
|
|
|
|
| 84 |
|
| 85 |
def transcribe_with_features(self, audio_path: str, retry_count: int = 0) -> Dict:
|
| 86 |
"""Enhanced transcription with audio features for prosody analysis - Optimized with retry mechanism"""
|
src/apis/routes/speaking_route.py
CHANGED
|
@@ -511,7 +511,10 @@ async def assess_pronunciation(
|
|
| 511 |
await optimize_post_assessment_processing(result, reference_text)
|
| 512 |
|
| 513 |
# Add processing time
|
|
|
|
| 514 |
processing_time = time.time() - start_time
|
|
|
|
|
|
|
| 515 |
result["processing_info"]["processing_time"] = processing_time
|
| 516 |
|
| 517 |
# Convert numpy types for JSON serialization
|
|
|
|
| 511 |
await optimize_post_assessment_processing(result, reference_text)
|
| 512 |
|
| 513 |
# Add processing time
|
| 514 |
+
|
| 515 |
processing_time = time.time() - start_time
|
| 516 |
+
if "processing_info" not in result:
|
| 517 |
+
result["processing_info"] = {}
|
| 518 |
result["processing_info"]["processing_time"] = processing_time
|
| 519 |
|
| 520 |
# Convert numpy types for JSON serialization
|