Spaces:
Sleeping
Sleeping
Updated some stuffs
Browse files- requirements.txt +1 -0
- src/model/predict.py +20 -7
- src/model/transcriber.py +24 -0
- src/predictions/feedback.csv +1 -0
- views/application.py +39 -8
- views/real_time.py +284 -9
requirements.txt
CHANGED
|
@@ -15,3 +15,4 @@ scikit-learn
|
|
| 15 |
huggingface
|
| 16 |
huggingface_hub
|
| 17 |
pyaudio
|
|
|
|
|
|
| 15 |
huggingface
|
| 16 |
huggingface_hub
|
| 17 |
pyaudio
|
| 18 |
+
streamlit_audiorec
|
src/model/predict.py
CHANGED
|
@@ -13,18 +13,31 @@ model.eval()
|
|
| 13 |
|
| 14 |
emotion_labels = ["joie", "colère", "neutre"]
|
| 15 |
|
| 16 |
-
def predict_emotion(audio_path):
|
| 17 |
-
waveform, _ = librosa.load(audio_path, sr=
|
| 18 |
-
input_values = processor(waveform, return_tensors="pt", sampling_rate=
|
| 19 |
input_values = input_values.to(device)
|
| 20 |
|
| 21 |
with torch.no_grad():
|
| 22 |
outputs = model(input_values)
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
| 27 |
# Exemple d'utilisation
|
| 28 |
audio_test = "data/n1ac.wav"
|
| 29 |
emotion = predict_emotion(audio_test)
|
| 30 |
-
print(f"Émotion détectée : {emotion}")
|
|
|
|
| 13 |
|
| 14 |
emotion_labels = ["joie", "colère", "neutre"]
|
| 15 |
|
| 16 |
+
def predict_emotion(audio_path, output_probs=False, sampling_rate=16000):
|
| 17 |
+
waveform, _ = librosa.load(audio_path, sr=sampling_rate)
|
| 18 |
+
input_values = processor(waveform, return_tensors="pt", sampling_rate=sampling_rate).input_values
|
| 19 |
input_values = input_values.to(device)
|
| 20 |
|
| 21 |
with torch.no_grad():
|
| 22 |
outputs = model(input_values)
|
| 23 |
+
|
| 24 |
+
if output_probs:
|
| 25 |
+
# Appliquer softmax pour obtenir des probabilités
|
| 26 |
+
probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)
|
| 27 |
+
|
| 28 |
+
# Convertir en numpy array et prendre le premier (et seul) élément
|
| 29 |
+
probabilities = probabilities[0].detach().cpu().numpy()
|
| 30 |
+
|
| 31 |
+
# Créer un dictionnaire associant chaque émotion à sa probabilité
|
| 32 |
+
emotion_probabilities = {emotion: prob for emotion, prob in zip(emotion_labels, probabilities)}
|
| 33 |
+
return emotion_probabilities
|
| 34 |
+
else:
|
| 35 |
+
# Obtenir l'émotion la plus probable (i.e. la prédiction)
|
| 36 |
+
predicted_label = torch.argmax(outputs, dim=1).item()
|
| 37 |
+
return emotion_labels[predicted_label]
|
| 38 |
+
|
| 39 |
|
| 40 |
# Exemple d'utilisation
|
| 41 |
audio_test = "data/n1ac.wav"
|
| 42 |
emotion = predict_emotion(audio_test)
|
| 43 |
+
print(f"Émotion détectée : {emotion}")
|
src/model/transcriber.py
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from transformers import Wav2Vec2Processor
|
| 3 |
+
from src.model.emotion_classifier import Wav2Vec2EmotionClassifier
|
| 4 |
+
import librosa
|
| 5 |
+
|
| 6 |
+
# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 7 |
+
# processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-large-xlsr-53")
|
| 8 |
+
# model = Wav2Vec2EmotionClassifier()
|
| 9 |
+
# model.load_state_dict(torch.load("wav2vec2_emotion.pth"))
|
| 10 |
+
# model.to(device)
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
def transcribe_audio(audio, sampling_rate=16000):
|
| 14 |
+
# Préparer les données d'entrée pour le modèle
|
| 15 |
+
input_values = processor(audio, sampling_rate=sampling_rate, return_tensors="pt").input_values
|
| 16 |
+
|
| 17 |
+
# Passer les données dans le modèle pour obtenir les logits
|
| 18 |
+
with torch.no_grad():
|
| 19 |
+
logits = model(input_values).logits
|
| 20 |
+
|
| 21 |
+
# Décoder les prédictions en texte
|
| 22 |
+
predicted_ids = torch.argmax(logits, dim=-1)
|
| 23 |
+
transcription = processor.batch_decode(predicted_ids)[0]
|
| 24 |
+
return transcription
|
src/predictions/feedback.csv
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
filepath,prediction,feedback
|
views/application.py
CHANGED
|
@@ -2,6 +2,7 @@ import streamlit as st
|
|
| 2 |
from st_audiorec import st_audiorec
|
| 3 |
import datetime
|
| 4 |
import os
|
|
|
|
| 5 |
|
| 6 |
DIRECTORY = "audios"
|
| 7 |
FILE_NAME = "audio.wav"
|
|
@@ -18,11 +19,11 @@ def application():
|
|
| 18 |
|
| 19 |
st.markdown("---")
|
| 20 |
|
| 21 |
-
tab1, tab2 = st.tabs(["⬆️ Record Audio", "🔈 Realtime Audio"])
|
| 22 |
|
| 23 |
with tab1:
|
| 24 |
-
st.header("⬆️ Upload Audio")
|
| 25 |
-
st.write("Here you can
|
| 26 |
audio_file = st.file_uploader("Upload an audio file", type=["wav", "mp3", "ogg"])
|
| 27 |
if audio_file is not None:
|
| 28 |
|
|
@@ -31,14 +32,44 @@ def application():
|
|
| 31 |
st.success(f"Saved file: {FILE_NAME}")
|
| 32 |
|
| 33 |
with tab2:
|
| 34 |
-
st.header("🔈 Realtime Audio")
|
| 35 |
-
st.write("Here you can
|
| 36 |
|
| 37 |
if st.button("Register", key="register-button"):
|
| 38 |
st.success("Audio registered successfully.")
|
| 39 |
|
| 40 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 41 |
|
| 42 |
-
if wav_audio_data is not None:
|
| 43 |
-
st.audio(wav_audio_data, format='audio/wav')
|
| 44 |
|
|
|
|
| 2 |
from st_audiorec import st_audiorec
|
| 3 |
import datetime
|
| 4 |
import os
|
| 5 |
+
from src.model.transcriber import transcribe_audio
|
| 6 |
|
| 7 |
DIRECTORY = "audios"
|
| 8 |
FILE_NAME = "audio.wav"
|
|
|
|
| 19 |
|
| 20 |
st.markdown("---")
|
| 21 |
|
| 22 |
+
tab1, tab2, tab3 = st.tabs(["⬆️ Record Audio", "🔈 Realtime Audio", "📝 Transcription"])
|
| 23 |
|
| 24 |
with tab1:
|
| 25 |
+
st.header("⬆️ Upload Audio Record")
|
| 26 |
+
st.write("Here you can upload a pre-recorded audio.")
|
| 27 |
audio_file = st.file_uploader("Upload an audio file", type=["wav", "mp3", "ogg"])
|
| 28 |
if audio_file is not None:
|
| 29 |
|
|
|
|
| 32 |
st.success(f"Saved file: {FILE_NAME}")
|
| 33 |
|
| 34 |
with tab2:
|
| 35 |
+
st.header("🔈 Realtime Audio Record")
|
| 36 |
+
st.write("Here you can record an audio.")
|
| 37 |
|
| 38 |
if st.button("Register", key="register-button"):
|
| 39 |
st.success("Audio registered successfully.")
|
| 40 |
|
| 41 |
+
audio_file = st_audiorec()
|
| 42 |
+
|
| 43 |
+
if audio_file is not None:
|
| 44 |
+
st.audio(audio_file, format='audio/wav')
|
| 45 |
+
|
| 46 |
+
with tab3:
|
| 47 |
+
st.header("📝 Speech2Text Transcription")
|
| 48 |
+
st.write("Here you can get the audio transcript.")
|
| 49 |
+
|
| 50 |
+
save = st.checkbox("Save transcription to .txt", value=False, key="save-transcript")
|
| 51 |
+
|
| 52 |
+
############################# A décommenté quand ce sera débogué
|
| 53 |
+
if st.button("Transcribe", key="transcribe-button"):
|
| 54 |
+
# # Fonction pour transcrire l'audio
|
| 55 |
+
# transcription = transcribe_audio(st.audio)
|
| 56 |
+
|
| 57 |
+
# # Charger et transcrire l'audio
|
| 58 |
+
# # audio, rate = load_audio(audio_file_path) # (re)chargement de l'audio si nécessaire
|
| 59 |
+
# transcription = transcribe_audio(audio_file, sampling_rate=16000)
|
| 60 |
+
|
| 61 |
+
# # Afficher la transcription
|
| 62 |
+
# st.write("Transcription :", transcription)
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
st.success("Audio registered successfully.")
|
| 66 |
+
# if save:
|
| 67 |
+
# file_path = "transcript.txt"
|
| 68 |
+
|
| 69 |
+
# # Write the text to the file
|
| 70 |
+
# with open(file_path, "w") as file:
|
| 71 |
+
# file.write(transcription)
|
| 72 |
+
|
| 73 |
+
# st.success(f"Text saved to {file_path}")
|
| 74 |
|
|
|
|
|
|
|
| 75 |
|
views/real_time.py
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
################################
|
| 2 |
-
###
|
| 3 |
###############################
|
| 4 |
|
| 5 |
import streamlit as st
|
|
@@ -7,8 +7,10 @@ import pyaudio
|
|
| 7 |
import wave
|
| 8 |
import torch
|
| 9 |
from transformers import Wav2Vec2ForSequenceClassification, Wav2Vec2Processor
|
|
|
|
| 10 |
import numpy as np
|
| 11 |
import time
|
|
|
|
| 12 |
|
| 13 |
# Charger le modèle Wav2Vec2 pour la classification des émotions
|
| 14 |
model_name = "superb/wav2vec2-base-superb-er" # Exemple de modèle pour la reconnaissance des émotions
|
|
@@ -22,13 +24,13 @@ CHANNELS = 1
|
|
| 22 |
RATE = 16000
|
| 23 |
|
| 24 |
# Fonction pour prédire l'émotion à partir d'un segment audio
|
| 25 |
-
def predict_emotion(audio_data):
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
|
| 33 |
# Interface Streamlit
|
| 34 |
st.title("Détection des émotions en temps réel")
|
|
@@ -56,7 +58,7 @@ if start_button:
|
|
| 56 |
# Traitement en temps réel (par tranche de 1 seconde)
|
| 57 |
if len(frames) >= RATE // CHUNK:
|
| 58 |
audio_segment = np.frombuffer(b''.join(frames[-(RATE // CHUNK):]), dtype=np.int16)
|
| 59 |
-
emotion = predict_emotion(audio_segment)
|
| 60 |
real_time_emotions.append(emotion)
|
| 61 |
emotion_placeholder.line_chart(real_time_emotions) # Affichage graphique des émotions
|
| 62 |
|
|
@@ -78,3 +80,276 @@ if start_button:
|
|
| 78 |
final_emotion = predict_emotion(full_audio_data)
|
| 79 |
|
| 80 |
final_emotion_placeholder.write(f"Émotion finale prédite : {final_emotion}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
################################
|
| 2 |
+
### Real time prediction for real time record
|
| 3 |
###############################
|
| 4 |
|
| 5 |
import streamlit as st
|
|
|
|
| 7 |
import wave
|
| 8 |
import torch
|
| 9 |
from transformers import Wav2Vec2ForSequenceClassification, Wav2Vec2Processor
|
| 10 |
+
import matplotlib.pyplot as plt
|
| 11 |
import numpy as np
|
| 12 |
import time
|
| 13 |
+
from src.model.predict import predict_emotion
|
| 14 |
|
| 15 |
# Charger le modèle Wav2Vec2 pour la classification des émotions
|
| 16 |
model_name = "superb/wav2vec2-base-superb-er" # Exemple de modèle pour la reconnaissance des émotions
|
|
|
|
| 24 |
RATE = 16000
|
| 25 |
|
| 26 |
# Fonction pour prédire l'émotion à partir d'un segment audio
|
| 27 |
+
# def predict_emotion(audio_data):
|
| 28 |
+
# inputs = processor(audio_data, sampling_rate=RATE, return_tensors="pt", padding=True)
|
| 29 |
+
# with torch.no_grad():
|
| 30 |
+
# logits = model(**inputs).logits
|
| 31 |
+
# predicted_id = torch.argmax(logits, dim=-1).item()
|
| 32 |
+
# emotion = model.config.id2label[predicted_id]
|
| 33 |
+
# return emotion
|
| 34 |
|
| 35 |
# Interface Streamlit
|
| 36 |
st.title("Détection des émotions en temps réel")
|
|
|
|
| 58 |
# Traitement en temps réel (par tranche de 1 seconde)
|
| 59 |
if len(frames) >= RATE // CHUNK:
|
| 60 |
audio_segment = np.frombuffer(b''.join(frames[-(RATE // CHUNK):]), dtype=np.int16)
|
| 61 |
+
emotion = predict_emotion(audio_segment, output_probs=False, sampling_rate=RATE)
|
| 62 |
real_time_emotions.append(emotion)
|
| 63 |
emotion_placeholder.line_chart(real_time_emotions) # Affichage graphique des émotions
|
| 64 |
|
|
|
|
| 80 |
final_emotion = predict_emotion(full_audio_data)
|
| 81 |
|
| 82 |
final_emotion_placeholder.write(f"Émotion finale prédite : {final_emotion}")
|
| 83 |
+
|
| 84 |
+
|
| 85 |
+
################################
|
| 86 |
+
### Real time prediction for uploaded audio file
|
| 87 |
+
###############################
|
| 88 |
+
# Charger le modèle wav2vec et le processeur
|
| 89 |
+
model = Wav2Vec2ForSequenceClassification.from_pretrained("your_emotion_model_path")
|
| 90 |
+
processor = Wav2Vec2Processor.from_pretrained("your_emotion_model_path")
|
| 91 |
+
|
| 92 |
+
# Définir les émotions
|
| 93 |
+
emotions = ["neutre", "joie", "colère", "tristesse"] # Ajustez selon votre modèle
|
| 94 |
+
|
| 95 |
+
# Fonction pour prédire l'émotion
|
| 96 |
+
# def predict_emotion(audio_chunk):
|
| 97 |
+
# inputs = processor(audio_chunk, sampling_rate=16000, return_tensors="pt", padding=True)
|
| 98 |
+
# with torch.no_grad():
|
| 99 |
+
# logits = model(**inputs).logits
|
| 100 |
+
# scores = torch.softmax(logits, dim=1).squeeze().tolist()
|
| 101 |
+
# return dict(zip(emotions, scores))
|
| 102 |
+
|
| 103 |
+
# Configuration Streamlit
|
| 104 |
+
st.title("Analyse des émotions en temps réel")
|
| 105 |
+
uploaded_file = st.file_uploader("Choisissez un fichier audio", type=["wav", "mp3"])
|
| 106 |
+
|
| 107 |
+
if uploaded_file is not None:
|
| 108 |
+
# Charger et rééchantillonner l'audio
|
| 109 |
+
audio, sr = librosa.load(uploaded_file, sr=16000)
|
| 110 |
+
|
| 111 |
+
# Paramètres de la fenêtre glissante
|
| 112 |
+
window_size = 1 # en secondes
|
| 113 |
+
hop_length = 0.5 # en secondes
|
| 114 |
+
|
| 115 |
+
# Créer un graphique en temps réel
|
| 116 |
+
fig, ax = plt.subplots()
|
| 117 |
+
lines = [ax.plot([], [], label=emotion)[0] for emotion in emotions]
|
| 118 |
+
ax.set_ylim(0, 1)
|
| 119 |
+
ax.set_xlim(0, len(audio) / sr)
|
| 120 |
+
ax.set_xlabel("Temps (s)")
|
| 121 |
+
ax.set_ylabel("Probabilité")
|
| 122 |
+
ax.legend()
|
| 123 |
+
|
| 124 |
+
chart = st.pyplot(fig)
|
| 125 |
+
|
| 126 |
+
# Traitement par fenêtre glissante
|
| 127 |
+
for i in range(0, len(audio), int(hop_length * sr)):
|
| 128 |
+
chunk = audio[i:i + int(window_size * sr)]
|
| 129 |
+
if len(chunk) < int(window_size * sr):
|
| 130 |
+
break
|
| 131 |
+
|
| 132 |
+
emotion_scores = predict_emotion(chunk, output_probs=False, sampling_rate=RATE)
|
| 133 |
+
|
| 134 |
+
# Mettre à jour le graphique
|
| 135 |
+
for emotion, line in zip(emotions, lines):
|
| 136 |
+
xdata = line.get_xdata().tolist()
|
| 137 |
+
ydata = line.get_ydata().tolist()
|
| 138 |
+
xdata.append(i / sr)
|
| 139 |
+
ydata.append(emotion_scores[emotion])
|
| 140 |
+
line.set_data(xdata, ydata)
|
| 141 |
+
|
| 142 |
+
ax.relim()
|
| 143 |
+
ax.autoscale_view()
|
| 144 |
+
chart.pyplot(fig)
|
| 145 |
+
|
| 146 |
+
st.success("Analyse terminée !")
|
| 147 |
+
|
| 148 |
+
|
| 149 |
+
|
| 150 |
+
|
| 151 |
+
|
| 152 |
+
|
| 153 |
+
|
| 154 |
+
############################################
|
| 155 |
+
### Progress bar
|
| 156 |
+
############################################
|
| 157 |
+
|
| 158 |
+
with st.status("Downloading data...", expanded=True) as status:
|
| 159 |
+
st.write("Searching for data...")
|
| 160 |
+
time.sleep(2)
|
| 161 |
+
st.write("Found URL.")
|
| 162 |
+
time.sleep(1)
|
| 163 |
+
st.write("Downloading data...")
|
| 164 |
+
time.sleep(1)
|
| 165 |
+
status.update(
|
| 166 |
+
label="Download complete!", state="complete", expanded=False
|
| 167 |
+
)
|
| 168 |
+
|
| 169 |
+
st.button("Rerun")
|
| 170 |
+
|
| 171 |
+
|
| 172 |
+
############################################
|
| 173 |
+
### Time duration estimation
|
| 174 |
+
############################################
|
| 175 |
+
progress_bar = st.progress(0)
|
| 176 |
+
time_placeholder = st.empty()
|
| 177 |
+
|
| 178 |
+
total_time = 10 # Total estimated time in seconds
|
| 179 |
+
for i in range(total_time):
|
| 180 |
+
# Update progress bar
|
| 181 |
+
progress_bar.progress((i + 1) / total_time)
|
| 182 |
+
|
| 183 |
+
# Update time estimation
|
| 184 |
+
remaining_time = total_time - i - 1
|
| 185 |
+
time_placeholder.text(f"Estimated time remaining: {remaining_time} seconds")
|
| 186 |
+
|
| 187 |
+
# Simulate task progress
|
| 188 |
+
time.sleep(1)
|
| 189 |
+
|
| 190 |
+
|
| 191 |
+
|
| 192 |
+
############################################
|
| 193 |
+
### Audio file noise reduction
|
| 194 |
+
############################################
|
| 195 |
+
from pydub import AudioSegment
|
| 196 |
+
import noisereduce as nr
|
| 197 |
+
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
| 198 |
+
|
| 199 |
+
# Fonction de réduction de bruit
|
| 200 |
+
def reduce_noise(audio_data, sr):
|
| 201 |
+
reduced_noise = nr.reduce_noise(y=audio_data, sr=sr)
|
| 202 |
+
return reduced_noise
|
| 203 |
+
|
| 204 |
+
# Chargement du modèle wav2vec
|
| 205 |
+
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
|
| 206 |
+
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")
|
| 207 |
+
|
| 208 |
+
# Interface Streamlit
|
| 209 |
+
st.title("Application de transcription audio avec réduction de bruit")
|
| 210 |
+
|
| 211 |
+
uploaded_file = st.file_uploader("Choisissez un fichier audio .wav", type="wav")
|
| 212 |
+
|
| 213 |
+
if uploaded_file is not None:
|
| 214 |
+
# Chargement et prétraitement de l'audio
|
| 215 |
+
audio = AudioSegment.from_wav(uploaded_file)
|
| 216 |
+
audio_array = np.array(audio.get_array_of_samples())
|
| 217 |
+
|
| 218 |
+
# Réduction de bruit
|
| 219 |
+
reduced_noise_audio = reduce_noise(audio_array, audio.frame_rate)
|
| 220 |
+
|
| 221 |
+
# Traitement avec wav2vec
|
| 222 |
+
input_values = processor(reduced_noise_audio, sampling_rate=audio.frame_rate, return_tensors="pt").input_values
|
| 223 |
+
|
| 224 |
+
with torch.no_grad():
|
| 225 |
+
logits = model(input_values).logits
|
| 226 |
+
|
| 227 |
+
predicted_ids = torch.argmax(logits, dim=-1)
|
| 228 |
+
transcription = processor.batch_decode(predicted_ids)[0]
|
| 229 |
+
|
| 230 |
+
st.audio(uploaded_file, format="audio/wav")
|
| 231 |
+
st.write("Transcription:")
|
| 232 |
+
st.write(transcription)
|
| 233 |
+
|
| 234 |
+
|
| 235 |
+
############################################
|
| 236 |
+
### Choix des émotions
|
| 237 |
+
############################################
|
| 238 |
+
# options = ['Sadness','Anger', 'Disgust', 'Fear', 'Surprise', 'Joy','Neutral']
|
| 239 |
+
# selected_options = st.multiselect('What emotions do you want to be displayed', options, default=['Joy', 'Anger','Neutral])
|
| 240 |
+
|
| 241 |
+
|
| 242 |
+
############################################
|
| 243 |
+
### Transcription Speech2Text
|
| 244 |
+
############################################
|
| 245 |
+
# # Fonction pour transcrire l'audio
|
| 246 |
+
# def transcribe_audio(audio):
|
| 247 |
+
# # Préparer les données d'entrée pour le modèle
|
| 248 |
+
# input_values = processor(audio, sampling_rate=16000, return_tensors="pt").input_values
|
| 249 |
+
|
| 250 |
+
# # Passer les données dans le modèle pour obtenir les logits
|
| 251 |
+
# with torch.no_grad():
|
| 252 |
+
# logits = model(input_values).logits
|
| 253 |
+
|
| 254 |
+
# # Décoder les prédictions en texte
|
| 255 |
+
# predicted_ids = torch.argmax(logits, dim=-1)
|
| 256 |
+
# transcription = processor.batch_decode(predicted_ids)[0]
|
| 257 |
+
# return transcription
|
| 258 |
+
|
| 259 |
+
# # Charger et transcrire l'audio
|
| 260 |
+
# # audio, rate = load_audio(audio_file_path) # (re)chargement de l'audio si nécessaire
|
| 261 |
+
# transcription = transcribe_audio(audio)
|
| 262 |
+
|
| 263 |
+
# # Afficher la transcription
|
| 264 |
+
# print("Transcription :", transcription)
|
| 265 |
+
|
| 266 |
+
|
| 267 |
+
############################################
|
| 268 |
+
### Feedback
|
| 269 |
+
############################################
|
| 270 |
+
import pandas as pd
|
| 271 |
+
import os
|
| 272 |
+
|
| 273 |
+
# Initialisation du fichier CSV
|
| 274 |
+
csv_file = "predictions/feedback.csv"
|
| 275 |
+
|
| 276 |
+
# Vérifier si le fichier CSV existe, sinon le créer avec des colonnes appropriées
|
| 277 |
+
if not os.path.exists(csv_file):
|
| 278 |
+
df = pd.DataFrame(columns=["filepath", "prediction", "feedback"])
|
| 279 |
+
df.to_csv(csv_file, index=False)
|
| 280 |
+
|
| 281 |
+
# Charger les données existantes du CSV
|
| 282 |
+
df = pd.read_csv(csv_file)
|
| 283 |
+
|
| 284 |
+
# Interface Streamlit
|
| 285 |
+
st.title("Predicted emotion feedback")
|
| 286 |
+
|
| 287 |
+
# Simuler une prédiction pour l'exemple (remplacez par votre modèle réel)
|
| 288 |
+
audio_file_name = "example_audio.wav"
|
| 289 |
+
predicted_emotion = "Joie" # Exemple de prédiction
|
| 290 |
+
|
| 291 |
+
st.write(f"Fichier audio : {audio_file_name}")
|
| 292 |
+
st.write(f"Émotion détectée : {predicted_emotion}")
|
| 293 |
+
|
| 294 |
+
# Formulaire de feedback
|
| 295 |
+
with st.form("feedback_form"):
|
| 296 |
+
st.write("Est-ce la bonne émotion qui a été détectée ? Cochez la réelle émotion.")
|
| 297 |
+
feedback = st.selectbox("Votre réponse :", ['Sadness','Anger', 'Disgust', 'Fear', 'Surprise', 'Joy', 'Neutral'])
|
| 298 |
+
submit_button = st.form_submit_button("Soumettre")
|
| 299 |
+
st.write("En cliquant sur ce bouton, vous acceptez que votre audio soit sauvegardé dans notre base de données.")
|
| 300 |
+
|
| 301 |
+
if submit_button:
|
| 302 |
+
# Ajouter le feedback au DataFrame
|
| 303 |
+
new_entry = {"filepath": audio_file_name, "prediction": predicted_emotion, "feedback": feedback}
|
| 304 |
+
df = df.append(new_entry, ignore_index=True)
|
| 305 |
+
|
| 306 |
+
# Sauvegarder les données mises à jour dans le fichier CSV
|
| 307 |
+
df.to_csv(csv_file, index=False)
|
| 308 |
+
|
| 309 |
+
# Sauvegarder le fichier audio
|
| 310 |
+
with open("predictions/data", "wb") as f:
|
| 311 |
+
f.write(uploaded_file.getbuffer())
|
| 312 |
+
|
| 313 |
+
# Confirmation pour l'utilisateur
|
| 314 |
+
st.success("Merci pour votre retour ! Vos données ont été sauvegardées.")
|
| 315 |
+
|
| 316 |
+
# Afficher les données sauvegardées (optionnel)
|
| 317 |
+
# st.write("Données collectées jusqu'à présent :")
|
| 318 |
+
# st.dataframe(df)
|
| 319 |
+
|
| 320 |
+
|
| 321 |
+
|
| 322 |
+
|
| 323 |
+
|
| 324 |
+
|
| 325 |
+
|
| 326 |
+
|
| 327 |
+
|
| 328 |
+
|
| 329 |
+
|
| 330 |
+
|
| 331 |
+
|
| 332 |
+
|
| 333 |
+
|
| 334 |
+
############################################
|
| 335 |
+
### Predict proba (to replace in predict.py)
|
| 336 |
+
############################################
|
| 337 |
+
import librosa
|
| 338 |
+
def predict_emotion_probabilities(audio_path):
|
| 339 |
+
waveform, _ = librosa.load(audio_path, sr=16000)
|
| 340 |
+
input_values = processor(waveform, return_tensors="pt", sampling_rate=16000).input_values
|
| 341 |
+
input_values = input_values.to(device)
|
| 342 |
+
|
| 343 |
+
with torch.no_grad():
|
| 344 |
+
outputs = model(input_values)
|
| 345 |
+
|
| 346 |
+
# Appliquer softmax pour obtenir des probabilités
|
| 347 |
+
probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)
|
| 348 |
+
|
| 349 |
+
# Convertir en numpy array et prendre le premier (et seul) élément
|
| 350 |
+
probabilities = probabilities[0].detach().cpu().numpy()
|
| 351 |
+
|
| 352 |
+
# Créer un dictionnaire associant chaque émotion à sa probabilité
|
| 353 |
+
emotion_probabilities = {emotion: prob for emotion, prob in zip(emotion_labels, probabilities)}
|
| 354 |
+
|
| 355 |
+
return emotion_probabilities
|