Update app.py
Browse files
app.py
CHANGED
@@ -2,6 +2,7 @@ import gradio as gr
|
|
2 |
import torch
|
3 |
import numpy as np
|
4 |
from transformers import Wav2Vec2FeatureExtractor, Wav2Vec2ForSequenceClassification
|
|
|
5 |
|
6 |
# Initialize model and processor
|
7 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
@@ -21,58 +22,119 @@ EMOTION_LABELS = {
|
|
21 |
6: "surprise"
|
22 |
}
|
23 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
def process_audio(audio):
|
25 |
"""Process audio chunk and return emotion"""
|
26 |
if audio is None:
|
27 |
-
return ""
|
28 |
-
|
29 |
try:
|
30 |
# Get the audio data
|
31 |
if isinstance(audio, tuple):
|
32 |
audio = audio[1]
|
33 |
-
|
34 |
-
# Convert to numpy array and ensure float32 type
|
35 |
-
audio = np.array(audio, dtype=np.float32)
|
36 |
-
|
37 |
-
# Ensure we have mono audio
|
38 |
-
if len(audio.shape) > 1:
|
39 |
-
audio = audio.mean(axis=1)
|
40 |
-
|
41 |
-
# Normalize audio if needed
|
42 |
-
if audio.max() > 1.0 or audio.min() < -1.0:
|
43 |
-
audio = audio / max(abs(audio.max()), abs(audio.min()))
|
44 |
|
45 |
-
#
|
46 |
-
|
47 |
-
|
48 |
-
|
|
|
|
|
|
|
|
|
49 |
# Prepare input for the model
|
50 |
inputs = feature_extractor(
|
51 |
-
|
52 |
sampling_rate=16000,
|
53 |
return_tensors="pt",
|
54 |
padding=True
|
55 |
)
|
56 |
|
57 |
-
#
|
58 |
inputs = {k: v.to(device, dtype=torch.float32) for k, v in inputs.items()}
|
59 |
|
60 |
# Get prediction
|
61 |
with torch.no_grad():
|
62 |
outputs = model(**inputs)
|
63 |
logits = outputs.logits
|
64 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
65 |
|
66 |
-
|
67 |
-
probs = torch.nn.functional.softmax(logits, dim=-1)
|
68 |
-
confidence = probs[0][predicted_id].item() * 100
|
69 |
|
70 |
-
emotion = EMOTION_LABELS[predicted_id]
|
71 |
-
return f"{emotion} (confidence: {confidence:.1f}%)"
|
72 |
-
|
73 |
except Exception as e:
|
74 |
-
print(f"Error in
|
75 |
-
return "
|
76 |
|
77 |
# Create Gradio interface
|
78 |
demo = gr.Interface(
|
@@ -86,9 +148,12 @@ demo = gr.Interface(
|
|
86 |
show_label=True
|
87 |
)
|
88 |
],
|
89 |
-
outputs=gr.Textbox(
|
90 |
-
|
91 |
-
|
|
|
|
|
|
|
92 |
live=True,
|
93 |
allow_flagging=False
|
94 |
)
|
|
|
2 |
import torch
|
3 |
import numpy as np
|
4 |
from transformers import Wav2Vec2FeatureExtractor, Wav2Vec2ForSequenceClassification
|
5 |
+
import librosa
|
6 |
|
7 |
# Initialize model and processor
|
8 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
|
|
22 |
6: "surprise"
|
23 |
}
|
24 |
|
25 |
+
def preprocess_audio(audio, target_sr=16000):
|
26 |
+
"""Enhanced audio preprocessing"""
|
27 |
+
try:
|
28 |
+
# Convert to numpy array and ensure float32
|
29 |
+
audio = np.array(audio, dtype=np.float32)
|
30 |
+
|
31 |
+
# Convert to mono if stereo
|
32 |
+
if len(audio.shape) > 1:
|
33 |
+
audio = librosa.to_mono(audio.T)
|
34 |
+
|
35 |
+
# Resample if needed
|
36 |
+
if target_sr != 16000:
|
37 |
+
audio = librosa.resample(audio, orig_sr=target_sr, target_sr=16000)
|
38 |
+
|
39 |
+
# Apply preprocessing steps
|
40 |
+
# 1. Noise reduction
|
41 |
+
audio = librosa.effects.preemphasis(audio)
|
42 |
+
|
43 |
+
# 2. Normalize
|
44 |
+
audio = librosa.util.normalize(audio)
|
45 |
+
|
46 |
+
# 3. Voice activity detection
|
47 |
+
intervals = librosa.effects.split(audio, top_db=20)
|
48 |
+
if len(intervals) > 0:
|
49 |
+
audio = np.concatenate([audio[start:end] for start, end in intervals])
|
50 |
+
|
51 |
+
# 4. Ensure minimum length (1 second)
|
52 |
+
if len(audio) < 16000:
|
53 |
+
audio = np.pad(audio, (0, 16000 - len(audio)))
|
54 |
+
|
55 |
+
# 5. Take center 3 seconds if too long
|
56 |
+
if len(audio) > 48000: # 3 seconds at 16kHz
|
57 |
+
center = len(audio) // 2
|
58 |
+
start = center - 24000
|
59 |
+
end = center + 24000
|
60 |
+
audio = audio[start:end]
|
61 |
+
|
62 |
+
return audio
|
63 |
+
|
64 |
+
except Exception as e:
|
65 |
+
print(f"Preprocessing error: {str(e)}")
|
66 |
+
return None
|
67 |
+
|
68 |
+
def get_emotion_history():
|
69 |
+
"""Get emotion detection history"""
|
70 |
+
if not hasattr(get_emotion_history, "history"):
|
71 |
+
get_emotion_history.history = []
|
72 |
+
return get_emotion_history.history
|
73 |
+
|
74 |
def process_audio(audio):
|
75 |
"""Process audio chunk and return emotion"""
|
76 |
if audio is None:
|
77 |
+
return "No audio input detected"
|
78 |
+
|
79 |
try:
|
80 |
# Get the audio data
|
81 |
if isinstance(audio, tuple):
|
82 |
audio = audio[1]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
83 |
|
84 |
+
# Preprocess audio
|
85 |
+
processed_audio = preprocess_audio(audio)
|
86 |
+
if processed_audio is None:
|
87 |
+
return "Audio preprocessing failed"
|
88 |
+
|
89 |
+
if np.max(np.abs(processed_audio)) < 0.01:
|
90 |
+
return "Audio too quiet"
|
91 |
+
|
92 |
# Prepare input for the model
|
93 |
inputs = feature_extractor(
|
94 |
+
processed_audio,
|
95 |
sampling_rate=16000,
|
96 |
return_tensors="pt",
|
97 |
padding=True
|
98 |
)
|
99 |
|
100 |
+
# Move to device and ensure float32
|
101 |
inputs = {k: v.to(device, dtype=torch.float32) for k, v in inputs.items()}
|
102 |
|
103 |
# Get prediction
|
104 |
with torch.no_grad():
|
105 |
outputs = model(**inputs)
|
106 |
logits = outputs.logits
|
107 |
+
probs = torch.nn.functional.softmax(logits, dim=-1)[0]
|
108 |
+
|
109 |
+
# Get top 2 predictions
|
110 |
+
top2_probs, top2_ids = torch.topk(probs, 2)
|
111 |
+
|
112 |
+
# Convert to percentages
|
113 |
+
top2_probs = [p * 100 for p in top2_probs.cpu().numpy()]
|
114 |
+
top2_emotions = [EMOTION_LABELS[idx.item()] for idx in top2_ids]
|
115 |
+
|
116 |
+
# Update history
|
117 |
+
history = get_emotion_history()
|
118 |
+
history.append(top2_emotions[0])
|
119 |
+
if len(history) > 5:
|
120 |
+
history.pop(0)
|
121 |
+
|
122 |
+
# Get most common emotion in history
|
123 |
+
if len(history) >= 3:
|
124 |
+
from collections import Counter
|
125 |
+
most_common = Counter(history).most_common(1)[0][0]
|
126 |
+
else:
|
127 |
+
most_common = top2_emotions[0]
|
128 |
+
|
129 |
+
result = f"Primary: {top2_emotions[0]} ({top2_probs[0]:.1f}%)\n"
|
130 |
+
result += f"Secondary: {top2_emotions[1]} ({top2_probs[1]:.1f}%)\n"
|
131 |
+
result += f"Trending: {most_common}"
|
132 |
|
133 |
+
return result
|
|
|
|
|
134 |
|
|
|
|
|
|
|
135 |
except Exception as e:
|
136 |
+
print(f"Error in processing: {str(e)}")
|
137 |
+
return "Processing error. Please try again."
|
138 |
|
139 |
# Create Gradio interface
|
140 |
demo = gr.Interface(
|
|
|
148 |
show_label=True
|
149 |
)
|
150 |
],
|
151 |
+
outputs=gr.Textbox(
|
152 |
+
label="Detected Emotions",
|
153 |
+
lines=3
|
154 |
+
),
|
155 |
+
title="Enhanced Live Emotion Detection",
|
156 |
+
description="Speak naturally into your microphone. Shows primary and secondary emotions with confidence levels.",
|
157 |
live=True,
|
158 |
allow_flagging=False
|
159 |
)
|