Spaces:
Running
Running
Lundi
Browse files
app.py
CHANGED
@@ -2,44 +2,61 @@ import gradio as gr
|
|
2 |
import librosa
|
3 |
import numpy as np
|
4 |
import soundfile as sf
|
5 |
-
from transformers import pipeline
|
6 |
import os
|
|
|
|
|
|
|
7 |
|
8 |
-
|
|
|
9 |
stt_pipeline = pipeline("automatic-speech-recognition", model="boumehdi/wav2vec2-large-xlsr-moroccan-darija")
|
10 |
print("Modèle chargé avec succès !")
|
11 |
|
|
|
|
|
|
|
|
|
|
|
12 |
def reduce_noise(audio, sr):
|
13 |
-
"""Réduction du bruit
|
14 |
audio = librosa.effects.preemphasis(audio)
|
15 |
noise_threshold = np.percentile(np.abs(audio), 10)
|
16 |
audio = np.where(np.abs(audio) > noise_threshold, audio, 0)
|
17 |
return audio
|
18 |
|
19 |
-
def
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
"""
|
21 |
-
|
|
|
22 |
"""
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
subsegment_length = len(segment) // num_subsegments
|
36 |
-
for i in range(num_subsegments):
|
37 |
-
subseg = segment[i * subsegment_length: (i + 1) * subsegment_length]
|
38 |
-
segments.append(subseg)
|
39 |
-
else:
|
40 |
-
segments.append(segment)
|
41 |
-
|
42 |
-
return segments
|
43 |
|
44 |
def process_audio(audio_path):
|
45 |
print(f"Fichier reçu : {audio_path}")
|
@@ -52,23 +69,25 @@ def process_audio(audio_path):
|
|
52 |
# Réduction du bruit
|
53 |
audio = reduce_noise(audio, sr)
|
54 |
|
55 |
-
#
|
56 |
-
|
57 |
-
print(f"
|
|
|
|
|
|
|
58 |
|
59 |
-
# Transcrire chaque
|
60 |
result = []
|
61 |
-
for
|
62 |
-
temp_filename = f"
|
63 |
-
sf.write(temp_filename, np.array(
|
64 |
|
65 |
-
# Transcription du segment
|
66 |
transcription = stt_pipeline(temp_filename)
|
67 |
text = transcription["text"].strip()
|
68 |
|
69 |
-
# Vérification du contenu
|
70 |
if text:
|
71 |
-
result.append(f"
|
72 |
|
73 |
# Supprimer le fichier temporaire
|
74 |
os.remove(temp_filename)
|
@@ -88,8 +107,8 @@ iface = gr.Interface(
|
|
88 |
fn=process_audio,
|
89 |
inputs=gr.Audio(type="filepath"),
|
90 |
outputs="text",
|
91 |
-
title="Transcription
|
92 |
-
description="Upload un fichier audio pour transcription
|
93 |
)
|
94 |
|
95 |
iface.launch()
|
|
|
2 |
import librosa
|
3 |
import numpy as np
|
4 |
import soundfile as sf
|
|
|
5 |
import os
|
6 |
+
from transformers import pipeline
|
7 |
+
import torchaudio
|
8 |
+
from pyannote.audio import Pipeline
|
9 |
|
10 |
+
# Charger le modèle de reconnaissance vocale
|
11 |
+
print("Chargement du modèle Wav2Vec2...")
|
12 |
stt_pipeline = pipeline("automatic-speech-recognition", model="boumehdi/wav2vec2-large-xlsr-moroccan-darija")
|
13 |
print("Modèle chargé avec succès !")
|
14 |
|
15 |
+
# Charger le pipeline de diarisation (détection des speakers)
|
16 |
+
print("Chargement du modèle de diarisation...")
|
17 |
+
diarization_pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization", use_auth_token="YOUR_HF_TOKEN")
|
18 |
+
print("Modèle de diarisation chargé !")
|
19 |
+
|
20 |
def reduce_noise(audio, sr):
|
21 |
+
"""Réduction du bruit pour améliorer la transcription"""
|
22 |
audio = librosa.effects.preemphasis(audio)
|
23 |
noise_threshold = np.percentile(np.abs(audio), 10)
|
24 |
audio = np.where(np.abs(audio) > noise_threshold, audio, 0)
|
25 |
return audio
|
26 |
|
27 |
+
def diarize_audio(audio_path):
|
28 |
+
"""
|
29 |
+
Diarisation de l'audio : détecte qui parle et à quel moment.
|
30 |
+
Retourne une liste de (speaker, début, fin).
|
31 |
+
"""
|
32 |
+
diarization = diarization_pipeline(audio_path)
|
33 |
+
speaker_segments = {}
|
34 |
+
|
35 |
+
for turn, _, speaker in diarization.itertracks(yield_label=True):
|
36 |
+
start, end = turn.start, turn.end
|
37 |
+
if speaker not in speaker_segments:
|
38 |
+
speaker_segments[speaker] = []
|
39 |
+
speaker_segments[speaker].append((start, end))
|
40 |
+
|
41 |
+
return speaker_segments
|
42 |
+
|
43 |
+
def merge_speaker_segments(audio, sr, speaker_segments):
|
44 |
"""
|
45 |
+
Fusionne les segments d’un même speaker pour améliorer la précision.
|
46 |
+
Retourne un dictionnaire {speaker: signal_audio_fusionné}.
|
47 |
"""
|
48 |
+
merged_audio = {}
|
49 |
+
|
50 |
+
for speaker, segments in speaker_segments.items():
|
51 |
+
combined_audio = np.array([])
|
52 |
+
for start, end in segments:
|
53 |
+
start_sample = int(start * sr)
|
54 |
+
end_sample = int(end * sr)
|
55 |
+
combined_audio = np.concatenate((combined_audio, audio[start_sample:end_sample]))
|
56 |
+
|
57 |
+
merged_audio[speaker] = combined_audio
|
58 |
+
|
59 |
+
return merged_audio
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
|
61 |
def process_audio(audio_path):
|
62 |
print(f"Fichier reçu : {audio_path}")
|
|
|
69 |
# Réduction du bruit
|
70 |
audio = reduce_noise(audio, sr)
|
71 |
|
72 |
+
# Étape de diarisation : détection des speakers
|
73 |
+
speaker_segments = diarize_audio(audio_path)
|
74 |
+
print(f"Speakers détectés : {list(speaker_segments.keys())}")
|
75 |
+
|
76 |
+
# Fusionner les segments de chaque speaker
|
77 |
+
merged_audio = merge_speaker_segments(audio, sr, speaker_segments)
|
78 |
|
79 |
+
# Transcrire chaque speaker
|
80 |
result = []
|
81 |
+
for speaker, audio_data in merged_audio.items():
|
82 |
+
temp_filename = f"temp_{speaker}.wav"
|
83 |
+
sf.write(temp_filename, np.array(audio_data), sr)
|
84 |
|
85 |
+
# Transcription du segment fusionné
|
86 |
transcription = stt_pipeline(temp_filename)
|
87 |
text = transcription["text"].strip()
|
88 |
|
|
|
89 |
if text:
|
90 |
+
result.append(f"{speaker}: {text}")
|
91 |
|
92 |
# Supprimer le fichier temporaire
|
93 |
os.remove(temp_filename)
|
|
|
107 |
fn=process_audio,
|
108 |
inputs=gr.Audio(type="filepath"),
|
109 |
outputs="text",
|
110 |
+
title="Transcription avec Diarisation",
|
111 |
+
description="Upload un fichier audio pour une transcription avec détection des speakers."
|
112 |
)
|
113 |
|
114 |
iface.launch()
|