File size: 15,321 Bytes
3cf77dc
 
 
 
6d401a4
f6d1ff0
6d401a4
3cf77dc
 
6d401a4
 
 
f6d1ff0
 
 
3cf77dc
 
 
 
 
 
6d401a4
 
 
 
3cf77dc
 
f6d1ff0
3cf77dc
 
f6d1ff0
9464f08
3cf77dc
f6d1ff0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9464f08
f6d1ff0
 
9464f08
f6d1ff0
9464f08
 
f6d1ff0
9464f08
 
 
 
 
 
 
 
 
 
 
 
 
f6d1ff0
 
 
 
 
 
3cf77dc
f6d1ff0
6d401a4
 
 
 
3cf77dc
f6d1ff0
3cf77dc
f6d1ff0
 
 
 
 
 
 
 
9464f08
 
 
 
f6d1ff0
3cf77dc
f6d1ff0
 
3cf77dc
f6d1ff0
3cf77dc
 
6d401a4
 
 
 
3cf77dc
 
 
f6d1ff0
 
3cf77dc
 
 
 
 
 
 
f6d1ff0
6d401a4
 
 
f6d1ff0
 
6d401a4
 
f6d1ff0
6d401a4
 
 
 
3cf77dc
6d401a4
f6d1ff0
3cf77dc
f6d1ff0
6d401a4
 
 
f6d1ff0
6d401a4
 
 
f6d1ff0
 
6d401a4
 
f6d1ff0
6d401a4
f6d1ff0
 
 
 
 
9464f08
f6d1ff0
6d401a4
f6d1ff0
6d401a4
f6d1ff0
 
 
 
6d401a4
f6d1ff0
 
 
6d401a4
f6d1ff0
 
 
 
9464f08
6d401a4
f6d1ff0
 
 
 
 
 
 
9464f08
6d401a4
f6d1ff0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3cf77dc
9464f08
f6d1ff0
9464f08
 
 
 
f6d1ff0
 
 
 
 
6d401a4
f6d1ff0
 
6d401a4
 
 
 
 
 
f6d1ff0
6d401a4
 
f6d1ff0
 
 
 
 
 
 
 
 
 
 
6d401a4
f6d1ff0
9464f08
f6d1ff0
 
 
 
6d401a4
 
f6d1ff0
 
 
 
3cf77dc
 
 
9464f08
 
 
 
f6d1ff0
9464f08
 
f6d1ff0
6d401a4
f6d1ff0
 
 
 
 
 
 
9464f08
f6d1ff0
3cf77dc
 
48302e7
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
import os
import streamlit as st
import tempfile
import torch
import transformers
from transformers import pipeline, AutoModelForSequenceClassification, AutoTokenizer, Wav2Vec2Processor, Wav2Vec2ForSequenceClassification
import plotly.express as px
import logging
import warnings
import whisper
from pydub import AudioSegment
import time
import numpy as np
import librosa
import subprocess

# Suppress warnings for a clean console
logging.getLogger("torch").setLevel(logging.CRITICAL)
logging.getLogger("transformers").setLevel(logging.CRITICAL)
warnings.filterwarnings("ignore")
os.environ["TOKENIZERS_PARALLELISM"] = "false"

# Check if CUDA is available, otherwise use CPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

# Set Streamlit app layout
st.set_page_config(layout="wide", page_title="Advanced Voice Emotion Analyzer")

# Interface design
st.title("πŸŽ™οΈ Advanced Voice Emotion Analyzer")
st.write("Analyze all 27 emotions from uploaded audio with enhanced detection to avoid neutral defaults.")

# Audio Preprocessing
def make_audio_scarier(audio_path, output_path):
    try:
        commands = [
            f"ffmpeg -i {audio_path} -af 'asetrate=44100*0.8,aresample=44100' temp1.wav",
            f"ffmpeg -i temp1.wav -af 'reverb=0.8:0.2:0.5:0.5:0.5:0.5' temp2.wav",
            f"ffmpeg -i temp2.wav -af 'atempo=1.2' {output_path}"
        ]
        for cmd in commands:
            subprocess.run(cmd, shell=True, check=True)
        for temp_file in ["temp1.wav", "temp2.wav"]:
            if os.path.exists(temp_file):
                os.remove(temp_file)
    except Exception as e:
        st.error(f"Audio processing failed: {str(e)}")
        raise

# Audio Feature Extraction
def extract_audio_features(audio_path):
    try:
        y, sr = librosa.load(audio_path, sr=16000)
        pitch_mean = np.mean(librosa.piptrack(y=y, sr=sr)[0][librosa.piptrack(y=y, sr=sr)[0] > 0]) if np.any(librosa.piptrack(y=y, sr=sr)[0] > 0) else 0
        energy_mean = np.mean(librosa.feature.rms(y=y))
        zcr_mean = np.mean(librosa.feature.zero_crossing_rate(y))
        return {"pitch_mean": pitch_mean, "energy_mean": energy_mean, "zcr_mean": zcr_mean}
    except Exception as e:
        st.error(f"Audio feature extraction failed: {str(e)}")
        return {}

# Audio Emotion Classification with Wav2Vec2
@st.cache_resource
def get_audio_emotion_classifier():
    processor = Wav2Vec2Processor.from_pretrained("superb/wav2vec2-base-superb-er")
    model = Wav2Vec2ForSequenceClassification.from_pretrained("superb/wav2vec2-base-superb-er")
    model = model.to(device)
    return processor, model

def perform_audio_emotion_detection(audio_path):
    try:
        processor, model = get_audio_emotion_classifier()
        waveform, sample_rate = librosa.load(audio_path, sr=16000)
        inputs = processor(waveform, sampling_rate=16000, return_tensors="pt", padding=True)
        inputs = {k: v.to(device) for k, v in inputs.items()}
        with torch.no_grad():
            logits = model(**inputs).logits
        scores = torch.softmax(logits, dim=1).detach().cpu().numpy()[0]
        audio_emotions = ["neutral", "happy", "sad", "angry", "fearful", "surprise", "disgust"]
        emotion_dict = {emotion: float(scores[i]) for i, emotion in enumerate(audio_emotions)}
        top_emotion = audio_emotions[np.argmax(scores)]
        # Enhanced boosting based on audio features
        features = extract_audio_features(audio_path)
        if features.get("pitch_mean", 0) < 200 and features.get("energy_mean", 0) > 0.1 and features.get("zcr_mean", 0) > 0.1:
            emotion_dict["fearful"] = min(1.0, emotion_dict.get("fearful", 0) + 0.4)  # Increased boost
            top_emotion = "fearful" if emotion_dict["fearful"] > emotion_dict[top_emotion] else top_emotion
        elif features.get("energy_mean", 0) > 0.25:  # Stricter threshold
            emotion_dict["angry"] = min(1.0, emotion_dict.get("angry", 0) + 0.35)
            top_emotion = "angry" if emotion_dict["angry"] > emotion_dict[top_emotion] else top_emotion
        elif features.get("pitch_mean", 0) > 500 and features.get("energy_mean", 0) < 0.05:
            emotion_dict["sad"] = min(1.0, emotion_dict.get("sad", 0) + 0.3)
            top_emotion = "sad" if emotion_dict["sad"] > emotion_dict[top_emotion] else top_emotion
        elif features.get("energy_mean", 0) > 0.15 and features.get("pitch_mean", 0) > 300:
            emotion_dict["happy"] = min(1.0, emotion_dict.get("happy", 0) + 0.3)
            top_emotion = "happy" if emotion_dict["happy"] > emotion_dict[top_emotion] else top_emotion
        elif features.get("zcr_mean", 0) > 0.15 and features.get("energy_mean", 0) > 0.1:
            emotion_dict["surprise"] = min(1.0, emotion_dict.get("surprise", 0) + 0.25)
            top_emotion = "surprise" if emotion_dict["surprise"] > emotion_dict[top_emotion] else top_emotion
        # Fallback to avoid neutral if score is low
        if emotion_dict["neutral"] > 0.5 and max([v for k, v in emotion_dict.items() if k != "neutral"]) > 0.3:
            emotion_dict["neutral"] = max(0.0, emotion_dict["neutral"] - 0.2)  # Reduce neutral weight
            top_emotion = max(emotion_dict, key=emotion_dict.get)
        return emotion_dict, top_emotion
    except Exception as e:
        st.error(f"Audio emotion detection failed: {str(e)}")
        return {}, "unknown"

# Text Emotion Classification with RoBERTa
@st.cache_resource
def get_text_emotion_classifier():
    tokenizer = AutoTokenizer.from_pretrained("SamLowe/roberta-base-go_emotions", use_fast=True)
    model = AutoModelForSequenceClassification.from_pretrained("SamLowe/roberta-base-go_emotions")
    model = model.to(device)
    return pipeline("text-classification", model=model, tokenizer=tokenizer, top_k=None, device=-1 if device.type == "cpu" else 0)

def perform_text_emotion_detection(text):
    try:
        classifier = get_text_emotion_classifier()
        results = classifier(text)[0]
        emotions = ["admiration", "amusement", "anger", "annoyance", "approval", "caring", "confusion", 
                    "curiosity", "desire", "disappointment", "disapproval", "disgust", "embarrassment", 
                    "excitement", "fear", "gratitude", "grief", "joy", "love", "nervousness", "optimism", 
                    "pride", "realization", "relief", "remorse", "sadness", "surprise", "neutral"]
        emotions_dict = {result['label']: result['score'] for result in results if result['label'] in emotions}
        top_emotion = max(emotions_dict, key=emotions_dict.get)
        # Reduce neutral influence if other emotions are strong
        if emotions_dict.get("neutral", 0) > 0.5 and max([v for k, v in emotions_dict.items() if k != "neutral"]) > 0.4:
            emotions_dict["neutral"] = max(0.0, emotions_dict["neutral"] - 0.15)
            top_emotion = max(emotions_dict, key=emotions_dict.get)
        return emotions_dict, top_emotion
    except Exception as e:
        st.error(f"Text emotion detection failed: {str(e)}")
        return {}, "unknown"

# Sarcasm Detection
@st.cache_resource
def get_sarcasm_classifier():
    tokenizer = AutoTokenizer.from_pretrained("cardiffnlp/twitter-roberta-base-irony", use_fast=True)
    model = AutoModelForSequenceClassification.from_pretrained("cardiffnlp/twitter-roberta-base-irony")
    model = model.to(device)
    return pipeline("text-classification", model=model, tokenizer=tokenizer, device=-1 if device.type == "cpu" else 0)

def perform_sarcasm_detection(text):
    try:
        classifier = get_sarcasm_classifier()
        result = classifier(text)[0]
        is_sarcastic = result['label'] == "LABEL_1"
        sarcasm_score = result['score'] if is_sarcastic else 1 - result['score']
        return is_sarcastic, sarcasm_score
    except Exception as e:
        st.error(f"Sarcasm detection failed: {str(e)}")
        return False, 0.0

# Validate Audio
def validate_audio(audio_path):
    try:
        sound = AudioSegment.from_file(audio_path)
        if sound.dBFS < -50 or len(sound) < 1000:
            st.warning("Audio volume too low or too short. Please use a louder, longer audio.")
            return False
        return True
    except Exception:
        st.error("Invalid or corrupted audio file.")
        return False

# Speech Recognition with Whisper
@st.cache_resource
def load_whisper_model():
    return whisper.load_model("large-v3")

def transcribe_audio(audio_path):
    try:
        sound = AudioSegment.from_file(audio_path)
        temp_wav_path = os.path.join(tempfile.gettempdir(), "temp_converted.wav")
        sound = sound.set_frame_rate(16000).set_channels(1)
        sound.export(temp_wav_path, format="wav")
        model = load_whisper_model()
        result = model.transcribe(temp_wav_path, language="en")
        os.remove(temp_wav_path)
        return result["text"].strip()
    except Exception as e:
        st.error(f"Transcription failed: {str(e)}")
        return ""

# Process Audio Files
def process_audio_file(audio_data):
    temp_dir = tempfile.gettempdir()
    temp_file_path = os.path.join(temp_dir, f"audio_{int(time.time())}.wav")
    with open(temp_file_path, "wb") as f:
        f.write(audio_data.getvalue())
    if not validate_audio(temp_file_path):
        return None
    return temp_file_path

# Display Results
def display_analysis_results(audio_path):
    st.header("Audio Analysis")
    st.audio(audio_path)

    # Preprocess audio
    processed_audio_path = os.path.join(tempfile.gettempdir(), f"processed_{int(time.time())}.wav")
    make_audio_scarier(audio_path, processed_audio_path)

    # Audio emotion detection
    audio_emotions, audio_top_emotion = perform_audio_emotion_detection(processed_audio_path)
    st.subheader("Audio-Based Emotion")
    st.write(f"**Dominant Emotion:** {audio_top_emotion} (Score: {audio_emotions.get(audio_top_emotion, 0):.3f})")
    st.write("Audio Emotions:", audio_emotions)  # Debug output

    # Transcription and text emotion detection
    transcribed_text = transcribe_audio(processed_audio_path)
    st.subheader("Transcribed Text")
    st.text_area("Text", transcribed_text, height=100, disabled=True)
    if transcribed_text:
        text_emotions, text_top_emotion = perform_text_emotion_detection(transcribed_text)
        st.write(f"**Text-Based Dominant Emotion:** {text_top_emotion} (Score: {text_emotions.get(text_top_emotion, 0):.3f})")
        st.write("Text Emotions:", text_emotions)  # Debug output

    # Combine emotions (prioritize audio, map to 27 emotions)
    emotion_map = {
        "neutral": "neutral", "happy": "joy", "sad": "sadness", "angry": "anger", 
        "fearful": "fear", "surprise": "surprise", "disgust": "disgust"
    }
    combined_emotions = {emotion: 0 for emotion in ["admiration", "amusement", "anger", "annoyance", "approval", "caring", 
                                                   "confusion", "curiosity", "desire", "disappointment", "disapproval", 
                                                   "disgust", "embarrassment", "excitement", "fear", "gratitude", 
                                                   "grief", "joy", "love", "nervousness", "optimism", "pride", 
                                                   "realization", "relief", "remorse", "sadness", "surprise", "neutral"]}
    for audio_emotion, score in audio_emotions.items():
        mapped_emotion = emotion_map.get(audio_emotion, "neutral")
        combined_emotions[mapped_emotion] = max(combined_emotions[mapped_emotion], score * 0.7)
    if transcribed_text:
        for text_emotion, score in text_emotions.items():
            combined_emotions[text_emotion] = combined_emotions.get(text_emotion, 0) + score * 0.3

    # Avoid neutral if other emotions are competitive
    top_emotion = max(combined_emotions, key=combined_emotions.get)
    if combined_emotions["neutral"] > 0.5 and max([v for k, v in combined_emotions.items() if k != "neutral"]) > 0.4:
        combined_emotions["neutral"] = max(0.0, combined_emotions["neutral"] - 0.25)  # Stronger reduction
        top_emotion = max(combined_emotions, key=combined_emotions.get)

    sentiment = "POSITIVE" if top_emotion in ["admiration", "amusement", "approval", "caring", "desire", "excitement", 
                                             "gratitude", "joy", "love", "optimism", "pride", "relief"] else "NEGATIVE" if top_emotion in ["anger", "annoyance", "disappointment", "disapproval", "disgust", "embarrassment", "fear", "grief", "nervousness", "remorse", "sadness"] else "NEUTRAL"

    # Sarcasm detection
    is_sarcastic, sarcasm_score = perform_sarcasm_detection(transcribed_text) if transcribed_text else (False, 0.0)

    # Display results
    col1, col2 = st.columns([1, 2])
    with col1:
        st.subheader("Sentiment")
        sentiment_icon = "πŸ‘" if sentiment == "POSITIVE" else "πŸ‘Ž" if sentiment == "NEGATIVE" else "😐"
        st.markdown(f"**{sentiment_icon} {sentiment.capitalize()}** (Based on {top_emotion})")
        st.subheader("Sarcasm")
        sarcasm_icon = "😏" if is_sarcastic else "😐"
        st.markdown(f"**{sarcasm_icon} {'Detected' if is_sarcastic else 'Not Detected'}** (Score: {sarcasm_score:.3f})")

    with col2:
        st.subheader("Emotion Distribution")
        sorted_emotions = sorted(combined_emotions.items(), key=lambda x: x[1], reverse=True)[:10]
        emotions, scores = zip(*sorted_emotions)
        fig = px.bar(x=list(emotions), y=list(scores), labels={'x': 'Emotion', 'y': 'Score'},
                     title="Top Emotion Scores", color=list(emotions),
                     color_discrete_sequence=px.colors.qualitative.Bold)
        fig.update_layout(yaxis_range=[0, 1], showlegend=False, title_font_size=14)
        st.plotly_chart(fig, use_container_width=True)

    with st.expander("Details"):
        st.write(f"**Audio Features:** {extract_audio_features(processed_audio_path)}")
        st.write("""
        **How it works:**
        - Audio Emotion: Wav2Vec2 detects 7 emotions with feature-based boosts.
        - Transcription: Whisper converts audio to text.
        - Text Emotion: RoBERTa refines 27 emotions from text.
        - Sarcasm: Analyzes text for irony.
        **Accuracy depends on:** Audio quality, clarity, and noise.
        """)

    # Clean up
    for path in [audio_path, processed_audio_path]:
        if os.path.exists(path):
            os.remove(path)

# Main App Logic
def main():
    st.header("Upload Audio File")
    audio_file = st.file_uploader("Upload audio (wav, mp3, ogg)", type=["wav", "mp3", "ogg"])
    if audio_file:
        temp_audio_path = process_audio_file(audio_file)
        if temp_audio_path:
            if st.button("Analyze Audio"):
                with st.spinner("Analyzing..."):
                    display_analysis_results(temp_audio_path)

    st.sidebar.header("About")
    st.sidebar.write("""
    **Models Used:**
    - Audio: superb/wav2vec2-base-superb-er (7 emotions)
    - Text: SamLowe/roberta-base-go_emotions (27 emotions)
    - Sarcasm: cardiffnlp/twitter-roberta-base-irony
    - Speech: OpenAI Whisper (large-v3)
    **Note:** Recording is not supported on Hugging Face Spaces; use uploaded files.
    """)

if __name__ == "__main__":
    main()