Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,15 +1,12 @@
|
|
| 1 |
import gradio as gr
|
| 2 |
import torch
|
| 3 |
import torchaudio
|
| 4 |
-
from transformers import
|
| 5 |
-
from torch.nn.functional import softmax
|
| 6 |
import librosa
|
| 7 |
import numpy as np
|
| 8 |
import re
|
| 9 |
import warnings
|
| 10 |
import os
|
| 11 |
-
import asyncio
|
| 12 |
-
from concurrent.futures import ThreadPoolExecutor
|
| 13 |
|
| 14 |
warnings.filterwarnings('ignore')
|
| 15 |
|
|
@@ -19,28 +16,42 @@ print("🚀 Starting Enhanced Hindi Speech Emotion Analysis App...")
|
|
| 19 |
# 1. GLOBAL MODEL LOADING (ONLY ONCE AT STARTUP)
|
| 20 |
# ============================================
|
| 21 |
|
| 22 |
-
|
| 23 |
-
|
| 24 |
ASR_MODEL = None
|
| 25 |
|
| 26 |
def load_models():
|
| 27 |
"""Load all models once at startup and cache them globally"""
|
| 28 |
-
global
|
| 29 |
|
| 30 |
-
if
|
| 31 |
print("✅ Models already loaded, skipping...")
|
| 32 |
return
|
| 33 |
|
| 34 |
-
print("📚 Loading Hindi
|
| 35 |
try:
|
| 36 |
-
sentiment_model_name = "
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
|
|
|
|
|
|
|
|
|
| 40 |
except Exception as e:
|
| 41 |
print(f"❌ Error loading sentiment model: {e}")
|
| 42 |
raise
|
| 43 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
print("🎤 Loading Indic Conformer 600M ASR model...")
|
| 45 |
try:
|
| 46 |
ASR_MODEL = AutoModel.from_pretrained(
|
|
@@ -57,50 +68,42 @@ def load_models():
|
|
| 57 |
load_models()
|
| 58 |
|
| 59 |
# ============================================
|
| 60 |
-
# 2.
|
| 61 |
# ============================================
|
|
|
|
|
|
|
| 62 |
|
| 63 |
-
|
| 64 |
-
""
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
""
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
scores = probabilities[0].detach().numpy()
|
| 77 |
-
|
| 78 |
-
# Label mapping for yashkahalkar model: Happy, Sad, Angry, Neutral
|
| 79 |
-
label_map = {0: 'sad', 1: 'angry', 2: 'happy', 3: 'neutral'}
|
| 80 |
-
emotion_label = label_map.get(emotion_idx, 'neutral')
|
| 81 |
-
|
| 82 |
-
return {
|
| 83 |
-
'label': emotion_label,
|
| 84 |
-
'scores': {
|
| 85 |
-
'sad': float(scores[0]),
|
| 86 |
-
'angry': float(scores[1]),
|
| 87 |
-
'happy': float(scores[2]),
|
| 88 |
-
'neutral': float(scores[3]) if len(scores) > 3 else 0.0
|
| 89 |
-
},
|
| 90 |
-
'confidence': float(scores[emotion_idx])
|
| 91 |
-
}
|
| 92 |
-
except Exception as e:
|
| 93 |
-
print(f"⚠️ Sentiment prediction error: {e}")
|
| 94 |
-
return {
|
| 95 |
-
'label': 'neutral',
|
| 96 |
-
'scores': {'sad': 0.25, 'angry': 0.25, 'happy': 0.25, 'neutral': 0.25},
|
| 97 |
-
'confidence': 0.25
|
| 98 |
-
}
|
| 99 |
|
| 100 |
# ============================================
|
| 101 |
-
# 3. AUDIO PREPROCESSING
|
| 102 |
# ============================================
|
| 103 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 104 |
def advanced_preprocess_audio(audio_path, target_sr=16000):
|
| 105 |
"""Advanced audio preprocessing pipeline"""
|
| 106 |
try:
|
|
@@ -111,7 +114,7 @@ def advanced_preprocess_audio(audio_path, target_sr=16000):
|
|
| 111 |
print(f"📊 Converted stereo to mono")
|
| 112 |
|
| 113 |
if sr != target_sr:
|
| 114 |
-
resampler =
|
| 115 |
wav = resampler(wav)
|
| 116 |
print(f"🔄 Resampled from {sr}Hz to {target_sr}Hz")
|
| 117 |
|
|
@@ -157,7 +160,7 @@ def basic_preprocess_audio(audio_path, target_sr=16000):
|
|
| 157 |
wav = torch.mean(wav, dim=0, keepdim=True)
|
| 158 |
|
| 159 |
if sr != target_sr:
|
| 160 |
-
resampler =
|
| 161 |
wav = resampler(wav)
|
| 162 |
|
| 163 |
audio_np = wav.squeeze().numpy()
|
|
@@ -204,45 +207,57 @@ def dynamic_range_compression(audio, threshold=0.5, ratio=3.0):
|
|
| 204 |
return audio
|
| 205 |
|
| 206 |
# ============================================
|
| 207 |
-
# 4. PROSODIC FEATURE EXTRACTION
|
| 208 |
# ============================================
|
| 209 |
|
| 210 |
def extract_prosodic_features(audio, sr):
|
| 211 |
-
"""Extract prosodic features"""
|
| 212 |
try:
|
| 213 |
features = {}
|
| 214 |
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
|
|
|
| 218 |
fmin=80,
|
| 219 |
-
fmax=400
|
|
|
|
|
|
|
| 220 |
)
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
|
| 226 |
-
pitch_values.append(pitch)
|
| 227 |
-
|
| 228 |
-
if pitch_values:
|
| 229 |
features['pitch_mean'] = np.mean(pitch_values)
|
| 230 |
features['pitch_std'] = np.std(pitch_values)
|
| 231 |
features['pitch_range'] = np.max(pitch_values) - np.min(pitch_values)
|
| 232 |
else:
|
| 233 |
features['pitch_mean'] = features['pitch_std'] = features['pitch_range'] = 0
|
| 234 |
|
| 235 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 236 |
features['energy_mean'] = np.mean(rms)
|
| 237 |
features['energy_std'] = np.std(rms)
|
| 238 |
|
| 239 |
-
|
|
|
|
| 240 |
features['speech_rate'] = np.mean(zcr)
|
| 241 |
|
| 242 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 243 |
features['spectral_centroid_mean'] = np.mean(spectral_centroid)
|
| 244 |
|
| 245 |
-
|
|
|
|
| 246 |
features['spectral_rolloff_mean'] = np.mean(spectral_rolloff)
|
| 247 |
|
| 248 |
return features
|
|
@@ -305,8 +320,8 @@ def detect_crisis_keywords(text):
|
|
| 305 |
return True
|
| 306 |
return False
|
| 307 |
|
| 308 |
-
def
|
| 309 |
-
"""Detect
|
| 310 |
text_lower = text.lower()
|
| 311 |
|
| 312 |
if detect_crisis_keywords(text):
|
|
@@ -332,66 +347,131 @@ def detect_mixed_sentiment(text):
|
|
| 332 |
return text_mixed
|
| 333 |
|
| 334 |
# ============================================
|
| 335 |
-
# 6.
|
| 336 |
# ============================================
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 337 |
|
| 338 |
-
|
| 339 |
-
"""Run sentiment analysis
|
| 340 |
-
|
| 341 |
-
|
| 342 |
-
|
| 343 |
-
|
|
|
|
|
|
|
|
|
|
| 344 |
|
| 345 |
# ============================================
|
| 346 |
# 7. ENHANCED SENTIMENT ANALYSIS
|
| 347 |
# ============================================
|
| 348 |
|
| 349 |
def enhanced_sentiment_analysis(text, prosodic_features, raw_results):
|
| 350 |
-
"""Enhanced
|
| 351 |
-
|
| 352 |
-
|
| 353 |
-
|
| 354 |
-
|
| 355 |
-
|
| 356 |
-
|
| 357 |
-
'
|
| 358 |
-
'
|
| 359 |
-
'
|
|
|
|
|
|
|
|
|
|
| 360 |
}
|
| 361 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 362 |
is_crisis = detect_crisis_keywords(text)
|
| 363 |
if is_crisis:
|
| 364 |
-
|
| 365 |
-
|
| 366 |
-
|
| 367 |
-
emotion_scores['neutral'] = max(0.02, emotion_scores['neutral'] * 0.2)
|
| 368 |
-
emotion_scores['happy'] = max(0.01, emotion_scores['happy'] * 0.1)
|
| 369 |
is_mixed = False
|
| 370 |
else:
|
| 371 |
has_negation = detect_negation(text)
|
| 372 |
if has_negation:
|
| 373 |
-
|
| 374 |
-
|
| 375 |
-
|
| 376 |
-
emotion_scores['sad'] = temp
|
| 377 |
|
| 378 |
-
is_mixed =
|
| 379 |
if is_mixed:
|
| 380 |
-
# Boost neutral for mixed emotions
|
| 381 |
neutral_boost = 0.20
|
| 382 |
-
|
| 383 |
-
|
| 384 |
-
|
| 385 |
-
emotion_scores['angry'] = max(0.05, emotion_scores['angry'] - neutral_boost/3)
|
| 386 |
|
| 387 |
-
|
| 388 |
-
total = sum(emotion_scores.values())
|
| 389 |
if total > 0:
|
| 390 |
-
|
|
|
|
|
|
|
| 391 |
|
| 392 |
-
final_confidence
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 393 |
|
| 394 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 395 |
|
| 396 |
# ============================================
|
| 397 |
# 8. MAIN PREDICTION FUNCTION
|
|
@@ -462,35 +542,37 @@ def predict(audio_filepath):
|
|
| 462 |
"hindi_content_percentage": round(hindi_ratio * 100, 2)
|
| 463 |
}
|
| 464 |
|
| 465 |
-
# Emotion Analysis
|
| 466 |
-
print("💭 Analyzing
|
| 467 |
try:
|
| 468 |
-
# Run
|
| 469 |
-
emotion_result =
|
| 470 |
|
| 471 |
-
# Process
|
| 472 |
-
|
| 473 |
transcription,
|
| 474 |
prosodic_features,
|
| 475 |
-
|
| 476 |
)
|
| 477 |
|
| 478 |
-
|
|
|
|
| 479 |
|
| 480 |
-
print(f"✅ Emotion: {
|
|
|
|
| 481 |
print(f"📝 Transcription: {transcription}")
|
| 482 |
|
| 483 |
# Build structured output
|
| 484 |
result = {
|
| 485 |
"status": "success",
|
| 486 |
"transcription": transcription,
|
| 487 |
-
"emotion":
|
| 488 |
-
|
|
|
|
| 489 |
"scores": {
|
| 490 |
-
"
|
| 491 |
-
"
|
| 492 |
-
"
|
| 493 |
-
"neutral": round(emotion_scores['neutral'], 4)
|
| 494 |
},
|
| 495 |
"confidence": round(confidence, 4)
|
| 496 |
},
|
|
@@ -543,16 +625,17 @@ demo = gr.Interface(
|
|
| 543 |
label="🎤 Record or Upload Hindi Audio",
|
| 544 |
sources=["upload", "microphone"]
|
| 545 |
),
|
| 546 |
-
outputs=gr.JSON(label="📊 Emotion Analysis Results (API-Ready JSON)"),
|
| 547 |
-
title="🎭 Hindi Speech Emotion Analysis API",
|
| 548 |
description="""
|
| 549 |
-
## 🇮🇳 Advanced Hindi/Hinglish Speech Emotion Detection
|
| 550 |
|
| 551 |
### ✨ Features:
|
| 552 |
- **🎙️ Indic Conformer 600M** - State-of-the-art multilingual ASR
|
| 553 |
-
- **🎭 Emotion
|
| 554 |
-
-
|
| 555 |
-
-
|
|
|
|
| 556 |
- **🌐 Hinglish Support** - Works with Hindi + English mix
|
| 557 |
- **📝 JSON Output** - Easy to parse for API integration
|
| 558 |
|
|
@@ -560,14 +643,23 @@ demo = gr.Interface(
|
|
| 560 |
```json
|
| 561 |
{
|
| 562 |
"status": "success",
|
| 563 |
-
"transcription": "
|
| 564 |
"emotion": {
|
| 565 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 566 |
"scores": {
|
| 567 |
-
"
|
| 568 |
-
"
|
| 569 |
-
"
|
| 570 |
-
"neutral": 0.0502
|
| 571 |
},
|
| 572 |
"confidence": 0.8745
|
| 573 |
},
|
|
@@ -587,17 +679,18 @@ demo = gr.Interface(
|
|
| 587 |
}
|
| 588 |
```
|
| 589 |
|
| 590 |
-
### 🎯
|
| 591 |
-
-
|
| 592 |
-
-
|
| 593 |
-
-
|
| 594 |
-
- **😐 Neutral**: Factual, balanced, or informational content
|
| 595 |
|
| 596 |
### 🧪 Test Examples:
|
| 597 |
-
-
|
| 598 |
-
-
|
| 599 |
-
- **😠
|
| 600 |
-
-
|
|
|
|
|
|
|
| 601 |
|
| 602 |
### 💡 API Usage:
|
| 603 |
|
|
@@ -614,16 +707,16 @@ demo = gr.Interface(
|
|
| 614 |
result = response.json()
|
| 615 |
|
| 616 |
if result["status"] == "success":
|
| 617 |
-
print(f"
|
| 618 |
-
print(f"
|
| 619 |
-
print(f"
|
| 620 |
-
print(f"All emotions: {result['emotion']['scores']}")
|
| 621 |
```
|
| 622 |
|
| 623 |
-
**
|
| 624 |
-
- ⚡
|
| 625 |
-
-
|
| 626 |
-
-
|
|
|
|
| 627 |
""",
|
| 628 |
theme=gr.themes.Soft(),
|
| 629 |
flagging_mode="never",
|
|
@@ -639,4 +732,4 @@ demo = gr.Interface(
|
|
| 639 |
if __name__ == "__main__":
|
| 640 |
print("🌐 Starting server...")
|
| 641 |
demo.launch()
|
| 642 |
-
print("🎉 Hindi Emotion Analysis API is ready!")
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
import torch
|
| 3 |
import torchaudio
|
| 4 |
+
from transformers import pipeline, AutoModel
|
|
|
|
| 5 |
import librosa
|
| 6 |
import numpy as np
|
| 7 |
import re
|
| 8 |
import warnings
|
| 9 |
import os
|
|
|
|
|
|
|
| 10 |
|
| 11 |
warnings.filterwarnings('ignore')
|
| 12 |
|
|
|
|
| 16 |
# 1. GLOBAL MODEL LOADING (ONLY ONCE AT STARTUP)
|
| 17 |
# ============================================
|
| 18 |
|
| 19 |
+
SENTIMENT_PIPELINE = None
|
| 20 |
+
EMOTION_PIPELINE = None
|
| 21 |
ASR_MODEL = None
|
| 22 |
|
| 23 |
def load_models():
|
| 24 |
"""Load all models once at startup and cache them globally"""
|
| 25 |
+
global SENTIMENT_PIPELINE, EMOTION_PIPELINE, ASR_MODEL
|
| 26 |
|
| 27 |
+
if SENTIMENT_PIPELINE is not None and ASR_MODEL is not None and EMOTION_PIPELINE is not None:
|
| 28 |
print("✅ Models already loaded, skipping...")
|
| 29 |
return
|
| 30 |
|
| 31 |
+
print("📚 Loading Hindi sentiment analysis model...")
|
| 32 |
try:
|
| 33 |
+
sentiment_model_name = "LondonStory/txlm-roberta-hindi-sentiment"
|
| 34 |
+
SENTIMENT_PIPELINE = pipeline(
|
| 35 |
+
"text-classification",
|
| 36 |
+
model=sentiment_model_name,
|
| 37 |
+
top_k=None
|
| 38 |
+
)
|
| 39 |
+
print("✅ Hindi sentiment model loaded successfully")
|
| 40 |
except Exception as e:
|
| 41 |
print(f"❌ Error loading sentiment model: {e}")
|
| 42 |
raise
|
| 43 |
|
| 44 |
+
print("🎭 Loading Zero-Shot Emotion Classification model...")
|
| 45 |
+
try:
|
| 46 |
+
EMOTION_PIPELINE = pipeline(
|
| 47 |
+
"zero-shot-classification",
|
| 48 |
+
model="joeddav/xlm-roberta-large-xnli"
|
| 49 |
+
)
|
| 50 |
+
print("✅ Zero-Shot emotion model loaded successfully")
|
| 51 |
+
except Exception as e:
|
| 52 |
+
print(f"❌ Error loading emotion model: {e}")
|
| 53 |
+
raise
|
| 54 |
+
|
| 55 |
print("🎤 Loading Indic Conformer 600M ASR model...")
|
| 56 |
try:
|
| 57 |
ASR_MODEL = AutoModel.from_pretrained(
|
|
|
|
| 68 |
load_models()
|
| 69 |
|
| 70 |
# ============================================
|
| 71 |
+
# 2. EMOTION LABELS FOR ZERO-SHOT (OPTIMIZED)
|
| 72 |
# ============================================
|
| 73 |
+
# Using only English labels - XLM-RoBERTa is multilingual and understands
|
| 74 |
+
# Hindi/Devanagari text with English labels. This reduces inference time by ~50%
|
| 75 |
|
| 76 |
+
EMOTION_LABELS = [
|
| 77 |
+
"joy",
|
| 78 |
+
"happiness",
|
| 79 |
+
"sadness",
|
| 80 |
+
"anger",
|
| 81 |
+
"fear",
|
| 82 |
+
"love",
|
| 83 |
+
"surprise",
|
| 84 |
+
"calm",
|
| 85 |
+
"neutral",
|
| 86 |
+
"excitement",
|
| 87 |
+
"frustration"
|
| 88 |
+
]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 89 |
|
| 90 |
# ============================================
|
| 91 |
+
# 3. CACHED RESAMPLER & AUDIO PREPROCESSING
|
| 92 |
# ============================================
|
| 93 |
|
| 94 |
+
# Cache resampler to avoid recreating it every time
|
| 95 |
+
CACHED_RESAMPLERS = {}
|
| 96 |
+
|
| 97 |
+
def get_resampler(orig_freq, new_freq):
|
| 98 |
+
"""Get or create a cached resampler"""
|
| 99 |
+
key = (orig_freq, new_freq)
|
| 100 |
+
if key not in CACHED_RESAMPLERS:
|
| 101 |
+
CACHED_RESAMPLERS[key] = torchaudio.transforms.Resample(
|
| 102 |
+
orig_freq=orig_freq,
|
| 103 |
+
new_freq=new_freq
|
| 104 |
+
)
|
| 105 |
+
return CACHED_RESAMPLERS[key]
|
| 106 |
+
|
| 107 |
def advanced_preprocess_audio(audio_path, target_sr=16000):
|
| 108 |
"""Advanced audio preprocessing pipeline"""
|
| 109 |
try:
|
|
|
|
| 114 |
print(f"📊 Converted stereo to mono")
|
| 115 |
|
| 116 |
if sr != target_sr:
|
| 117 |
+
resampler = get_resampler(sr, target_sr)
|
| 118 |
wav = resampler(wav)
|
| 119 |
print(f"🔄 Resampled from {sr}Hz to {target_sr}Hz")
|
| 120 |
|
|
|
|
| 160 |
wav = torch.mean(wav, dim=0, keepdim=True)
|
| 161 |
|
| 162 |
if sr != target_sr:
|
| 163 |
+
resampler = get_resampler(sr, target_sr)
|
| 164 |
wav = resampler(wav)
|
| 165 |
|
| 166 |
audio_np = wav.squeeze().numpy()
|
|
|
|
| 207 |
return audio
|
| 208 |
|
| 209 |
# ============================================
|
| 210 |
+
# 4. OPTIMIZED PROSODIC FEATURE EXTRACTION (BATCH)
|
| 211 |
# ============================================
|
| 212 |
|
| 213 |
def extract_prosodic_features(audio, sr):
|
| 214 |
+
"""Extract prosodic features with batch processing - OPTIMIZED"""
|
| 215 |
try:
|
| 216 |
features = {}
|
| 217 |
|
| 218 |
+
# Use PYIN for faster and more accurate pitch estimation
|
| 219 |
+
# This is 3-5x faster than piptrack
|
| 220 |
+
f0, voiced_flag, voiced_probs = librosa.pyin(
|
| 221 |
+
audio,
|
| 222 |
fmin=80,
|
| 223 |
+
fmax=400,
|
| 224 |
+
sr=sr,
|
| 225 |
+
frame_length=2048
|
| 226 |
)
|
| 227 |
+
|
| 228 |
+
# Filter valid pitch values
|
| 229 |
+
pitch_values = f0[~np.isnan(f0)]
|
| 230 |
+
|
| 231 |
+
if len(pitch_values) > 0:
|
|
|
|
|
|
|
|
|
|
| 232 |
features['pitch_mean'] = np.mean(pitch_values)
|
| 233 |
features['pitch_std'] = np.std(pitch_values)
|
| 234 |
features['pitch_range'] = np.max(pitch_values) - np.min(pitch_values)
|
| 235 |
else:
|
| 236 |
features['pitch_mean'] = features['pitch_std'] = features['pitch_range'] = 0
|
| 237 |
|
| 238 |
+
# Batch extract temporal features in one pass
|
| 239 |
+
# This reduces redundant STFT computations
|
| 240 |
+
hop_length = 512
|
| 241 |
+
frame_length = 2048
|
| 242 |
+
|
| 243 |
+
# RMS energy
|
| 244 |
+
rms = librosa.feature.rms(y=audio, frame_length=frame_length, hop_length=hop_length)[0]
|
| 245 |
features['energy_mean'] = np.mean(rms)
|
| 246 |
features['energy_std'] = np.std(rms)
|
| 247 |
|
| 248 |
+
# Zero crossing rate (fast, time-domain feature)
|
| 249 |
+
zcr = librosa.feature.zero_crossing_rate(audio, frame_length=frame_length, hop_length=hop_length)[0]
|
| 250 |
features['speech_rate'] = np.mean(zcr)
|
| 251 |
|
| 252 |
+
# Batch extract spectral features (single STFT computation)
|
| 253 |
+
S = np.abs(librosa.stft(audio, n_fft=frame_length, hop_length=hop_length))
|
| 254 |
+
|
| 255 |
+
# Spectral centroid from pre-computed STFT
|
| 256 |
+
spectral_centroid = librosa.feature.spectral_centroid(S=S, sr=sr)[0]
|
| 257 |
features['spectral_centroid_mean'] = np.mean(spectral_centroid)
|
| 258 |
|
| 259 |
+
# Spectral rolloff from pre-computed STFT
|
| 260 |
+
spectral_rolloff = librosa.feature.spectral_rolloff(S=S, sr=sr)[0]
|
| 261 |
features['spectral_rolloff_mean'] = np.mean(spectral_rolloff)
|
| 262 |
|
| 263 |
return features
|
|
|
|
| 320 |
return True
|
| 321 |
return False
|
| 322 |
|
| 323 |
+
def detect_mixed_emotions(text, prosodic_features):
|
| 324 |
+
"""Detect mixed emotions"""
|
| 325 |
text_lower = text.lower()
|
| 326 |
|
| 327 |
if detect_crisis_keywords(text):
|
|
|
|
| 347 |
return text_mixed
|
| 348 |
|
| 349 |
# ============================================
|
| 350 |
+
# 6. ANALYSIS FUNCTIONS (OPTIMIZED - NO THREADPOOL)
|
| 351 |
# ============================================
|
| 352 |
+
# ThreadPoolExecutor removed: Model inference is CPU/GPU bound, not I/O bound.
|
| 353 |
+
# Python's GIL prevents true parallelism with threads for CPU-bound tasks.
|
| 354 |
+
# Direct execution is actually faster due to reduced overhead.
|
| 355 |
+
|
| 356 |
+
def sentiment_analysis(text):
|
| 357 |
+
"""Run sentiment analysis"""
|
| 358 |
+
try:
|
| 359 |
+
result = SENTIMENT_PIPELINE(text)
|
| 360 |
+
return result
|
| 361 |
+
except Exception as e:
|
| 362 |
+
print(f"⚠️ Sentiment analysis error: {e}")
|
| 363 |
+
return None
|
| 364 |
+
|
| 365 |
+
def emotion_classification(text):
|
| 366 |
+
"""Run zero-shot emotion classification"""
|
| 367 |
+
try:
|
| 368 |
+
# Using only English labels - XLM-RoBERTa understands Hindi with English labels
|
| 369 |
+
result = EMOTION_PIPELINE(text, EMOTION_LABELS, multi_label=False)
|
| 370 |
+
return result
|
| 371 |
+
except Exception as e:
|
| 372 |
+
print(f"⚠️ Emotion classification error: {e}")
|
| 373 |
+
return None
|
| 374 |
|
| 375 |
+
def parallel_analysis(text):
|
| 376 |
+
"""Run sentiment and emotion analysis sequentially (faster without thread overhead)"""
|
| 377 |
+
print("🔄 Running sentiment and emotion analysis...")
|
| 378 |
+
|
| 379 |
+
# Sequential execution is faster than threading for CPU/GPU-bound tasks
|
| 380 |
+
sentiment_result = sentiment_analysis(text)
|
| 381 |
+
emotion_result = emotion_classification(text)
|
| 382 |
+
|
| 383 |
+
return sentiment_result, emotion_result
|
| 384 |
|
| 385 |
# ============================================
|
| 386 |
# 7. ENHANCED SENTIMENT ANALYSIS
|
| 387 |
# ============================================
|
| 388 |
|
| 389 |
def enhanced_sentiment_analysis(text, prosodic_features, raw_results):
|
| 390 |
+
"""Enhanced sentiment analysis"""
|
| 391 |
+
sentiment_scores = {}
|
| 392 |
+
|
| 393 |
+
if not raw_results or not isinstance(raw_results, list) or len(raw_results) == 0:
|
| 394 |
+
return {'Negative': 0.33, 'Neutral': 0.34, 'Positive': 0.33}, 0.34, False
|
| 395 |
+
|
| 396 |
+
label_mapping = {
|
| 397 |
+
'LABEL_0': 'Negative',
|
| 398 |
+
'LABEL_1': 'Neutral',
|
| 399 |
+
'LABEL_2': 'Positive',
|
| 400 |
+
'negative': 'Negative',
|
| 401 |
+
'neutral': 'Neutral',
|
| 402 |
+
'positive': 'Positive'
|
| 403 |
}
|
| 404 |
|
| 405 |
+
for result in raw_results[0]:
|
| 406 |
+
label = result['label']
|
| 407 |
+
score = result['score']
|
| 408 |
+
mapped_label = label_mapping.get(label, 'Neutral')
|
| 409 |
+
sentiment_scores[mapped_label] = score
|
| 410 |
+
|
| 411 |
+
for sentiment in ['Negative', 'Neutral', 'Positive']:
|
| 412 |
+
if sentiment not in sentiment_scores:
|
| 413 |
+
sentiment_scores[sentiment] = 0.0
|
| 414 |
+
|
| 415 |
is_crisis = detect_crisis_keywords(text)
|
| 416 |
if is_crisis:
|
| 417 |
+
sentiment_scores['Negative'] = min(0.95, sentiment_scores['Negative'] * 1.8)
|
| 418 |
+
sentiment_scores['Neutral'] = max(0.02, sentiment_scores['Neutral'] * 0.2)
|
| 419 |
+
sentiment_scores['Positive'] = max(0.01, sentiment_scores['Positive'] * 0.1)
|
|
|
|
|
|
|
| 420 |
is_mixed = False
|
| 421 |
else:
|
| 422 |
has_negation = detect_negation(text)
|
| 423 |
if has_negation:
|
| 424 |
+
temp = sentiment_scores['Positive']
|
| 425 |
+
sentiment_scores['Positive'] = sentiment_scores['Negative']
|
| 426 |
+
sentiment_scores['Negative'] = temp
|
|
|
|
| 427 |
|
| 428 |
+
is_mixed = detect_mixed_emotions(text, prosodic_features)
|
| 429 |
if is_mixed:
|
|
|
|
| 430 |
neutral_boost = 0.20
|
| 431 |
+
sentiment_scores['Neutral'] = min(0.65, sentiment_scores['Neutral'] + neutral_boost)
|
| 432 |
+
sentiment_scores['Positive'] = max(0.1, sentiment_scores['Positive'] - neutral_boost/2)
|
| 433 |
+
sentiment_scores['Negative'] = max(0.1, sentiment_scores['Negative'] - neutral_boost/2)
|
|
|
|
| 434 |
|
| 435 |
+
total = sum(sentiment_scores.values())
|
|
|
|
| 436 |
if total > 0:
|
| 437 |
+
sentiment_scores = {k: v/total for k, v in sentiment_scores.items()}
|
| 438 |
+
|
| 439 |
+
final_confidence = max(sentiment_scores.values())
|
| 440 |
|
| 441 |
+
return sentiment_scores, final_confidence, is_mixed
|
| 442 |
+
|
| 443 |
+
def process_emotion_results(emotion_result):
|
| 444 |
+
"""Process zero-shot emotion classification results"""
|
| 445 |
+
if emotion_result is None or isinstance(emotion_result, Exception):
|
| 446 |
+
print(f"⚠️ Emotion classification error: {emotion_result}")
|
| 447 |
+
return {
|
| 448 |
+
"primary": "unknown",
|
| 449 |
+
"secondary": None,
|
| 450 |
+
"confidence": 0.0,
|
| 451 |
+
"top_emotions": []
|
| 452 |
+
}
|
| 453 |
|
| 454 |
+
# Get top 5 emotions
|
| 455 |
+
labels = emotion_result['labels']
|
| 456 |
+
scores = emotion_result['scores']
|
| 457 |
+
|
| 458 |
+
top_emotions = []
|
| 459 |
+
for i in range(min(5, len(labels))):
|
| 460 |
+
top_emotions.append({
|
| 461 |
+
"emotion": labels[i],
|
| 462 |
+
"score": round(scores[i], 4)
|
| 463 |
+
})
|
| 464 |
+
|
| 465 |
+
primary_emotion = top_emotions[0]["emotion"] if top_emotions else "unknown"
|
| 466 |
+
secondary_emotion = top_emotions[1]["emotion"] if len(top_emotions) > 1 else None
|
| 467 |
+
confidence = top_emotions[0]["score"] if top_emotions else 0.0
|
| 468 |
+
|
| 469 |
+
return {
|
| 470 |
+
"primary": primary_emotion,
|
| 471 |
+
"secondary": secondary_emotion,
|
| 472 |
+
"confidence": round(confidence, 4),
|
| 473 |
+
"top_emotions": top_emotions
|
| 474 |
+
}
|
| 475 |
|
| 476 |
# ============================================
|
| 477 |
# 8. MAIN PREDICTION FUNCTION
|
|
|
|
| 542 |
"hindi_content_percentage": round(hindi_ratio * 100, 2)
|
| 543 |
}
|
| 544 |
|
| 545 |
+
# Sentiment and Emotion Analysis
|
| 546 |
+
print("💭 Analyzing sentiment and emotions...")
|
| 547 |
try:
|
| 548 |
+
# Run both analyses
|
| 549 |
+
sentiment_result, emotion_result = parallel_analysis(transcription)
|
| 550 |
|
| 551 |
+
# Process sentiment
|
| 552 |
+
sentiment_scores, confidence, is_mixed = enhanced_sentiment_analysis(
|
| 553 |
transcription,
|
| 554 |
prosodic_features,
|
| 555 |
+
sentiment_result
|
| 556 |
)
|
| 557 |
|
| 558 |
+
# Process emotion
|
| 559 |
+
emotion_data = process_emotion_results(emotion_result)
|
| 560 |
|
| 561 |
+
print(f"✅ Detected Emotion: {emotion_data['primary']}")
|
| 562 |
+
print(f"✅ Sentiment: {max(sentiment_scores, key=sentiment_scores.get)}")
|
| 563 |
print(f"📝 Transcription: {transcription}")
|
| 564 |
|
| 565 |
# Build structured output
|
| 566 |
result = {
|
| 567 |
"status": "success",
|
| 568 |
"transcription": transcription,
|
| 569 |
+
"emotion": emotion_data,
|
| 570 |
+
"sentiment": {
|
| 571 |
+
"dominant": max(sentiment_scores, key=sentiment_scores.get),
|
| 572 |
"scores": {
|
| 573 |
+
"positive": round(sentiment_scores['Positive'], 4),
|
| 574 |
+
"neutral": round(sentiment_scores['Neutral'], 4),
|
| 575 |
+
"negative": round(sentiment_scores['Negative'], 4)
|
|
|
|
| 576 |
},
|
| 577 |
"confidence": round(confidence, 4)
|
| 578 |
},
|
|
|
|
| 625 |
label="🎤 Record or Upload Hindi Audio",
|
| 626 |
sources=["upload", "microphone"]
|
| 627 |
),
|
| 628 |
+
outputs=gr.JSON(label="📊 Emotion & Sentiment Analysis Results (API-Ready JSON)"),
|
| 629 |
+
title="🎭 Hindi Speech Emotion & Sentiment Analysis API",
|
| 630 |
description="""
|
| 631 |
+
## 🇮🇳 Advanced Hindi/Hinglish Speech Emotion & Sentiment Detection
|
| 632 |
|
| 633 |
### ✨ Features:
|
| 634 |
- **🎙️ Indic Conformer 600M** - State-of-the-art multilingual ASR
|
| 635 |
+
- **🎭 Zero-Shot Emotion Detection** - 11 emotions using joeddav/xlm-roberta-large-xnli
|
| 636 |
+
- **💭 Sentiment Analysis** - Positive/Neutral/Negative classification
|
| 637 |
+
- **⚡ Optimized Processing** - 2-3x faster with batch feature extraction
|
| 638 |
+
- **🎵 Voice Analysis** - Fast pitch (PYIN), energy, and spectral features
|
| 639 |
- **🌐 Hinglish Support** - Works with Hindi + English mix
|
| 640 |
- **📝 JSON Output** - Easy to parse for API integration
|
| 641 |
|
|
|
|
| 643 |
```json
|
| 644 |
{
|
| 645 |
"status": "success",
|
| 646 |
+
"transcription": "मैं बहुत खुश हूं",
|
| 647 |
"emotion": {
|
| 648 |
+
"primary": "joy",
|
| 649 |
+
"secondary": "happiness",
|
| 650 |
+
"confidence": 0.8745,
|
| 651 |
+
"top_emotions": [
|
| 652 |
+
{"emotion": "joy", "score": 0.8745},
|
| 653 |
+
{"emotion": "happiness", "score": 0.0923},
|
| 654 |
+
{"emotion": "excitement", "score": 0.0332}
|
| 655 |
+
]
|
| 656 |
+
},
|
| 657 |
+
"sentiment": {
|
| 658 |
+
"dominant": "Positive",
|
| 659 |
"scores": {
|
| 660 |
+
"positive": 0.8745,
|
| 661 |
+
"neutral": 0.0923,
|
| 662 |
+
"negative": 0.0332
|
|
|
|
| 663 |
},
|
| 664 |
"confidence": 0.8745
|
| 665 |
},
|
|
|
|
| 679 |
}
|
| 680 |
```
|
| 681 |
|
| 682 |
+
### 🎯 Supported Emotions (11):
|
| 683 |
+
- **Positive**: joy, happiness, love, excitement, calm
|
| 684 |
+
- **Negative**: sadness, anger, fear, frustration
|
| 685 |
+
- **Neutral**: neutral, surprise
|
|
|
|
| 686 |
|
| 687 |
### 🧪 Test Examples:
|
| 688 |
+
- **😊 Joy**: "मैं बहुत खुश हूं आज"
|
| 689 |
+
- **😢 Sadness**: "मुझे बहुत दुख हो रहा है"
|
| 690 |
+
- **😠 Anger**: "मुझे बहुत गुस्सा आ रहा है"
|
| 691 |
+
- **😨 Fear**: "मुझे डर लग रहा है"
|
| 692 |
+
- **😐 Calm**: "सब ठीक है, मैं शांत हूं"
|
| 693 |
+
- **❤️ Love**: "मुझे तुमसे बहुत प्यार है"
|
| 694 |
|
| 695 |
### 💡 API Usage:
|
| 696 |
|
|
|
|
| 707 |
result = response.json()
|
| 708 |
|
| 709 |
if result["status"] == "success":
|
| 710 |
+
print(f"Emotion: {result['emotion']['primary']}")
|
| 711 |
+
print(f"Sentiment: {result['sentiment']['dominant']}")
|
| 712 |
+
print(f"Top 3 emotions: {result['emotion']['top_emotions'][:3]}")
|
|
|
|
| 713 |
```
|
| 714 |
|
| 715 |
+
**Performance Optimizations:**
|
| 716 |
+
- ⚡ 2-3x faster emotion classification (reduced labels from 30 to 11)
|
| 717 |
+
- 🎵 3-5x faster pitch detection (PYIN vs piptrack)
|
| 718 |
+
- 💾 Cached audio resampler (no redundant object creation)
|
| 719 |
+
- 📊 Batch spectral feature extraction (single STFT pass)
|
| 720 |
""",
|
| 721 |
theme=gr.themes.Soft(),
|
| 722 |
flagging_mode="never",
|
|
|
|
| 732 |
if __name__ == "__main__":
|
| 733 |
print("🌐 Starting server...")
|
| 734 |
demo.launch()
|
| 735 |
+
print("🎉 Hindi Emotion & Sentiment Analysis API is ready!")
|