Commit
·
65b0afc
1
Parent(s):
700a4c7
Add preprocessing
Browse files- app.py +177 -10
- requirements.txt +1 -0
app.py
CHANGED
|
@@ -4,6 +4,7 @@ import tempfile
|
|
| 4 |
import time
|
| 5 |
import logging
|
| 6 |
import gc
|
|
|
|
| 7 |
from dataclasses import dataclass
|
| 8 |
from typing import Optional, Tuple, List, Any, Dict
|
| 9 |
from contextlib import contextmanager
|
|
@@ -12,11 +13,16 @@ import gradio as gr
|
|
| 12 |
import torch
|
| 13 |
import psutil
|
| 14 |
from dotenv import load_dotenv
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
|
| 16 |
load_dotenv()
|
| 17 |
|
| 18 |
-
# Audio preprocessing
|
| 19 |
-
PREPROCESSING_AVAILABLE =
|
| 20 |
|
| 21 |
|
| 22 |
def get_env_or_secret(key: str, default: Optional[str] = None) -> Optional[str]:
|
|
@@ -41,8 +47,143 @@ class PreprocessingConfig:
|
|
| 41 |
|
| 42 |
normalize_format: bool = True
|
| 43 |
normalize_volume: bool = True
|
| 44 |
-
reduce_noise: bool =
|
| 45 |
-
remove_silence: bool =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
|
| 47 |
|
| 48 |
def load_asr_pipeline(
|
|
@@ -208,6 +349,23 @@ def transcribe_local(
|
|
| 208 |
if not os.path.exists(audio_path):
|
| 209 |
raise FileNotFoundError(f"Audio file not found: {audio_path}")
|
| 210 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 211 |
# Load ASR pipeline with performance monitoring
|
| 212 |
start_time = time.time()
|
| 213 |
|
|
@@ -262,10 +420,10 @@ def transcribe_local(
|
|
| 262 |
try:
|
| 263 |
# Primary inference attempt with safe parameters
|
| 264 |
if asr_kwargs:
|
| 265 |
-
result = asr(
|
| 266 |
else:
|
| 267 |
# Fallback to no parameters if all failed
|
| 268 |
-
result = asr(
|
| 269 |
|
| 270 |
inference_time = time.time() - inference_start
|
| 271 |
memory_after = psutil.Process().memory_info().rss / 1024 / 1024 # MB
|
|
@@ -295,7 +453,7 @@ def transcribe_local(
|
|
| 295 |
|
| 296 |
try:
|
| 297 |
inference_start = time.time()
|
| 298 |
-
result = asr(
|
| 299 |
inference_time = time.time() - inference_start
|
| 300 |
memory_used = 0 # Reset memory tracking
|
| 301 |
|
|
@@ -313,6 +471,14 @@ def transcribe_local(
|
|
| 313 |
torch.cuda.empty_cache()
|
| 314 |
gc.collect()
|
| 315 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 316 |
# Return results with performance metrics
|
| 317 |
meta = {
|
| 318 |
"device": device_str,
|
|
@@ -320,6 +486,7 @@ def transcribe_local(
|
|
| 320 |
"inference_time": inference_time,
|
| 321 |
"memory_used_mb": memory_used,
|
| 322 |
"model_type": "original" if model_id == base_model_id else "fine-tuned",
|
|
|
|
| 323 |
}
|
| 324 |
|
| 325 |
return {"result": result, "meta": meta}
|
|
@@ -386,8 +553,8 @@ def transcribe_comparison(audio_file):
|
|
| 386 |
error_msg = "❌ Modelli non configurati. Impostare HF_MODEL_ID e BASE_WHISPER_MODEL_ID nelle variabili d'ambiente"
|
| 387 |
return error_msg, error_msg
|
| 388 |
|
| 389 |
-
# Preprocessing sempre attivo
|
| 390 |
-
#
|
| 391 |
|
| 392 |
# Fixed settings optimized for medical transcription
|
| 393 |
language = "it" # Always Italian for ScribeAId
|
|
@@ -540,7 +707,7 @@ def create_interface():
|
|
| 540 |
- Modello originale: `{base_model_id}`
|
| 541 |
- Modello fine-tuned: `{model_id}`
|
| 542 |
- Lingua: Italiano (it)
|
| 543 |
-
- Preprocessing audio:
|
| 544 |
""")
|
| 545 |
|
| 546 |
gr.Markdown("---")
|
|
|
|
| 4 |
import time
|
| 5 |
import logging
|
| 6 |
import gc
|
| 7 |
+
import io
|
| 8 |
from dataclasses import dataclass
|
| 9 |
from typing import Optional, Tuple, List, Any, Dict
|
| 10 |
from contextlib import contextmanager
|
|
|
|
| 13 |
import torch
|
| 14 |
import psutil
|
| 15 |
from dotenv import load_dotenv
|
| 16 |
+
import numpy as np
|
| 17 |
+
from pydub import AudioSegment
|
| 18 |
+
from pydub.silence import split_on_silence
|
| 19 |
+
import soundfile as sf
|
| 20 |
+
import noisereduce
|
| 21 |
|
| 22 |
load_dotenv()
|
| 23 |
|
| 24 |
+
# Audio preprocessing available with required dependencies
|
| 25 |
+
PREPROCESSING_AVAILABLE = True
|
| 26 |
|
| 27 |
|
| 28 |
def get_env_or_secret(key: str, default: Optional[str] = None) -> Optional[str]:
|
|
|
|
| 47 |
|
| 48 |
normalize_format: bool = True
|
| 49 |
normalize_volume: bool = True
|
| 50 |
+
reduce_noise: bool = True
|
| 51 |
+
remove_silence: bool = True
|
| 52 |
+
|
| 53 |
+
|
| 54 |
+
def normalize_audio(audio_bytes: bytes) -> bytes:
|
| 55 |
+
"""
|
| 56 |
+
Converte un chunk audio in bytes nel formato standard per Whisper.
|
| 57 |
+
(16kHz, mono, WAV PCM)
|
| 58 |
+
"""
|
| 59 |
+
# Carica i bytes in pydub usando un file in memoria (BytesIO)
|
| 60 |
+
audio_segment = AudioSegment.from_file(io.BytesIO(audio_bytes))
|
| 61 |
+
|
| 62 |
+
# 1. Imposta la frequenza di campionamento a 16kHz
|
| 63 |
+
audio_segment = audio_segment.set_frame_rate(16000)
|
| 64 |
+
# 2. Converte in mono
|
| 65 |
+
audio_segment = audio_segment.set_channels(1)
|
| 66 |
+
# 3. Assicura che il campione sia a 2 bytes (16-bit), standard per WAV
|
| 67 |
+
audio_segment = audio_segment.set_sample_width(2)
|
| 68 |
+
|
| 69 |
+
# Esporta i bytes processati in formato WAV
|
| 70 |
+
buffer = io.BytesIO()
|
| 71 |
+
audio_segment.export(buffer, format="wav")
|
| 72 |
+
return buffer.getvalue()
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
def normalize_volume(audio_bytes: bytes) -> bytes:
|
| 76 |
+
"""
|
| 77 |
+
Normalizza il volume di un chunk audio WAV.
|
| 78 |
+
"""
|
| 79 |
+
# Carica l'audio
|
| 80 |
+
audio_segment = AudioSegment.from_wav(io.BytesIO(audio_bytes))
|
| 81 |
+
|
| 82 |
+
# Normalizza l'audio. Porta il picco massimo a -1.0 dBFS
|
| 83 |
+
# Il valore di headroom è una buona pratica per evitare clipping
|
| 84 |
+
normalized_segment = audio_segment.normalize(headroom=0.1)
|
| 85 |
+
|
| 86 |
+
buffer = io.BytesIO()
|
| 87 |
+
normalized_segment.export(buffer, format="wav")
|
| 88 |
+
return buffer.getvalue()
|
| 89 |
+
|
| 90 |
+
|
| 91 |
+
def reduce_background_noise(audio_bytes: bytes) -> bytes:
|
| 92 |
+
"""
|
| 93 |
+
Riduce il rumore di fondo da un chunk audio WAV.
|
| 94 |
+
"""
|
| 95 |
+
# Leggi i dati audio dai bytes
|
| 96 |
+
buffer_read = io.BytesIO(audio_bytes)
|
| 97 |
+
rate, data = sf.read(buffer_read)
|
| 98 |
+
|
| 99 |
+
# Assicura che l'audio sia mono per la riduzione
|
| 100 |
+
if data.ndim > 1:
|
| 101 |
+
data = np.mean(data, axis=1)
|
| 102 |
+
|
| 103 |
+
# Esegui la riduzione del rumore
|
| 104 |
+
reduced_noise_data = noisereduce.reduce_noise(y=data, sr=rate)
|
| 105 |
+
|
| 106 |
+
# Scrivi i dati processati in un nuovo buffer di bytes
|
| 107 |
+
buffer_write = io.BytesIO()
|
| 108 |
+
sf.write(buffer_write, reduced_noise_data, rate, format="wav")
|
| 109 |
+
return buffer_write.getvalue()
|
| 110 |
+
|
| 111 |
+
|
| 112 |
+
def remove_silence(audio_bytes: bytes) -> bytes:
|
| 113 |
+
"""
|
| 114 |
+
Rimuove i segmenti di silenzio da un chunk audio in formato WAV.
|
| 115 |
+
"""
|
| 116 |
+
|
| 117 |
+
audio_segment = AudioSegment.from_wav(io.BytesIO(audio_bytes))
|
| 118 |
+
|
| 119 |
+
chunks = split_on_silence(
|
| 120 |
+
audio_segment,
|
| 121 |
+
min_silence_len=100,
|
| 122 |
+
silence_thresh=-35,
|
| 123 |
+
keep_silence=80, # Mantiene un piccolo silenzio tra i chunk
|
| 124 |
+
)
|
| 125 |
+
|
| 126 |
+
if not chunks:
|
| 127 |
+
# Se non trova parlato, restituisce bytes vuoti
|
| 128 |
+
return b""
|
| 129 |
+
|
| 130 |
+
# Unisce di nuovo i chunk in un unico segmento
|
| 131 |
+
processed_segment = sum(chunks, AudioSegment.empty())
|
| 132 |
+
|
| 133 |
+
buffer = io.BytesIO()
|
| 134 |
+
processed_segment.export(buffer, format="wav")
|
| 135 |
+
return buffer.getvalue()
|
| 136 |
+
|
| 137 |
+
|
| 138 |
+
def preprocess_audio_pipeline(audio_path: str) -> str:
|
| 139 |
+
"""
|
| 140 |
+
Applica la pipeline completa di preprocessing audio.
|
| 141 |
+
Restituisce il path del file audio preprocessato.
|
| 142 |
+
"""
|
| 143 |
+
logger = logging.getLogger(__name__)
|
| 144 |
+
logger.info("Avvio pipeline di preprocessing audio")
|
| 145 |
+
|
| 146 |
+
try:
|
| 147 |
+
# Leggi il file audio originale
|
| 148 |
+
with open(audio_path, "rb") as f:
|
| 149 |
+
audio_bytes = f.read()
|
| 150 |
+
|
| 151 |
+
# Applica tutte le fasi di preprocessing in sequenza
|
| 152 |
+
logger.info("1. Normalizzazione formato audio...")
|
| 153 |
+
audio_bytes = normalize_audio(audio_bytes)
|
| 154 |
+
|
| 155 |
+
logger.info("2. Normalizzazione volume...")
|
| 156 |
+
audio_bytes = normalize_volume(audio_bytes)
|
| 157 |
+
|
| 158 |
+
logger.info("3. Riduzione rumore di fondo...")
|
| 159 |
+
audio_bytes = reduce_background_noise(audio_bytes)
|
| 160 |
+
|
| 161 |
+
logger.info("4. Rimozione silenzi...")
|
| 162 |
+
audio_bytes = remove_silence(audio_bytes)
|
| 163 |
+
|
| 164 |
+
# Se l'audio è vuoto dopo la rimozione del silenzio, usa l'audio originale
|
| 165 |
+
if not audio_bytes:
|
| 166 |
+
logger.warning(
|
| 167 |
+
"Audio vuoto dopo rimozione silenzi, utilizzo audio originale"
|
| 168 |
+
)
|
| 169 |
+
with open(audio_path, "rb") as f:
|
| 170 |
+
audio_bytes = f.read()
|
| 171 |
+
# Applica solo normalizzazione formato e volume
|
| 172 |
+
audio_bytes = normalize_audio(audio_bytes)
|
| 173 |
+
audio_bytes = normalize_volume(audio_bytes)
|
| 174 |
+
|
| 175 |
+
# Salva l'audio preprocessato in un file temporaneo
|
| 176 |
+
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as temp_file:
|
| 177 |
+
temp_file.write(audio_bytes)
|
| 178 |
+
preprocessed_path = temp_file.name
|
| 179 |
+
|
| 180 |
+
logger.info(f"Preprocessing completato: {preprocessed_path}")
|
| 181 |
+
return preprocessed_path
|
| 182 |
+
|
| 183 |
+
except Exception as e:
|
| 184 |
+
logger.error(f"Errore durante preprocessing: {e}")
|
| 185 |
+
logger.info("Utilizzo audio originale senza preprocessing")
|
| 186 |
+
return audio_path
|
| 187 |
|
| 188 |
|
| 189 |
def load_asr_pipeline(
|
|
|
|
| 349 |
if not os.path.exists(audio_path):
|
| 350 |
raise FileNotFoundError(f"Audio file not found: {audio_path}")
|
| 351 |
|
| 352 |
+
# Apply audio preprocessing pipeline
|
| 353 |
+
preprocessed_audio_path = audio_path
|
| 354 |
+
if PREPROCESSING_AVAILABLE:
|
| 355 |
+
try:
|
| 356 |
+
logger.info("Applicazione preprocessing audio...")
|
| 357 |
+
preprocessed_audio_path = preprocess_audio_pipeline(audio_path)
|
| 358 |
+
logger.info(
|
| 359 |
+
f"Preprocessing completato. File processato: {os.path.basename(preprocessed_audio_path)}"
|
| 360 |
+
)
|
| 361 |
+
except Exception as e:
|
| 362 |
+
logger.warning(
|
| 363 |
+
f"Errore durante preprocessing, utilizzo audio originale: {e}"
|
| 364 |
+
)
|
| 365 |
+
preprocessed_audio_path = audio_path
|
| 366 |
+
else:
|
| 367 |
+
logger.info("Preprocessing audio non disponibile, utilizzo audio originale")
|
| 368 |
+
|
| 369 |
# Load ASR pipeline with performance monitoring
|
| 370 |
start_time = time.time()
|
| 371 |
|
|
|
|
| 420 |
try:
|
| 421 |
# Primary inference attempt with safe parameters
|
| 422 |
if asr_kwargs:
|
| 423 |
+
result = asr(preprocessed_audio_path, **asr_kwargs)
|
| 424 |
else:
|
| 425 |
# Fallback to no parameters if all failed
|
| 426 |
+
result = asr(preprocessed_audio_path)
|
| 427 |
|
| 428 |
inference_time = time.time() - inference_start
|
| 429 |
memory_after = psutil.Process().memory_info().rss / 1024 / 1024 # MB
|
|
|
|
| 453 |
|
| 454 |
try:
|
| 455 |
inference_start = time.time()
|
| 456 |
+
result = asr(preprocessed_audio_path) # No parameters at all
|
| 457 |
inference_time = time.time() - inference_start
|
| 458 |
memory_used = 0 # Reset memory tracking
|
| 459 |
|
|
|
|
| 471 |
torch.cuda.empty_cache()
|
| 472 |
gc.collect()
|
| 473 |
|
| 474 |
+
# Cleanup temporary preprocessed file if it was created
|
| 475 |
+
if preprocessed_audio_path != audio_path:
|
| 476 |
+
try:
|
| 477 |
+
os.unlink(preprocessed_audio_path)
|
| 478 |
+
logger.info("File audio preprocessato temporaneo rimosso")
|
| 479 |
+
except Exception as e:
|
| 480 |
+
logger.warning(f"Errore rimozione file temporaneo: {e}")
|
| 481 |
+
|
| 482 |
# Return results with performance metrics
|
| 483 |
meta = {
|
| 484 |
"device": device_str,
|
|
|
|
| 486 |
"inference_time": inference_time,
|
| 487 |
"memory_used_mb": memory_used,
|
| 488 |
"model_type": "original" if model_id == base_model_id else "fine-tuned",
|
| 489 |
+
"preprocessing_applied": preprocessed_audio_path != audio_path,
|
| 490 |
}
|
| 491 |
|
| 492 |
return {"result": result, "meta": meta}
|
|
|
|
| 553 |
error_msg = "❌ Modelli non configurati. Impostare HF_MODEL_ID e BASE_WHISPER_MODEL_ID nelle variabili d'ambiente"
|
| 554 |
return error_msg, error_msg
|
| 555 |
|
| 556 |
+
# Preprocessing sempre attivo: normalizzazione formato, volume, riduzione rumore, rimozione silenzi
|
| 557 |
+
# Viene applicato automaticamente prima della trascrizione con entrambi i modelli
|
| 558 |
|
| 559 |
# Fixed settings optimized for medical transcription
|
| 560 |
language = "it" # Always Italian for ScribeAId
|
|
|
|
| 707 |
- Modello originale: `{base_model_id}`
|
| 708 |
- Modello fine-tuned: `{model_id}`
|
| 709 |
- Lingua: Italiano (it)
|
| 710 |
+
- Preprocessing audio: **ATTIVO** (normalizzazione, riduzione rumore, rimozione silenzi)
|
| 711 |
""")
|
| 712 |
|
| 713 |
gr.Markdown("---")
|
requirements.txt
CHANGED
|
@@ -12,3 +12,4 @@ psutil>=5.9.0
|
|
| 12 |
python-dotenv>=1.0.0
|
| 13 |
datasets>=2.14.0
|
| 14 |
huggingface-hub>=0.17.0
|
|
|
|
|
|
| 12 |
python-dotenv>=1.0.0
|
| 13 |
datasets>=2.14.0
|
| 14 |
huggingface-hub>=0.17.0
|
| 15 |
+
noisereduce>=3.0.0
|