app.py

import gradio as gr
import torch
import librosa
import numpy as np
from transformers import pipeline, AutoConfig
import gc
import warnings
import os
warnings.filterwarnings("ignore")

# Set environment variables for optimization
os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
os.environ["TOKENIZERS_PARALLELISM"] = "false"

class UltraLightShukaASR:
    def __init__(self):
        self.pipe = None
        self.model_loaded = False
        
    def load_model_lazy(self):
        """Lazy load model only when needed"""
        if self.model_loaded:
            return True
            
        try:
            print("Loading Shuka v1 model...")
            
            # Try with minimal configuration first
            self.pipe = pipeline(
                model='sarvamai/shuka-1',
                trust_remote_code=True,
                device=-1,  # CPU only
                model_kwargs={
                    "low_cpu_mem_usage": True,
                    "use_cache": False,  # Disable cache to save memory
                    "torch_dtype": torch.float32,
                }
            )
            
            print("✅ Model loaded successfully!")
            self.model_loaded = True
            return True
            
        except Exception as e:
            print(f"❌ Model loading failed: {e}")
            return False
    
    def preprocess_audio_minimal(self, audio_input, target_sr=16000, max_duration=15):
        """Minimal audio preprocessing for speed"""
        try:
            if isinstance(audio_input, tuple):
                sr, audio_data = audio_input
                audio_data = audio_data.astype(np.float32)
                if len(audio_data.shape) > 1:
                    audio_data = np.mean(audio_data, axis=1)
            else:
                audio_data, sr = librosa.load(audio_input, sr=target_sr, duration=max_duration)
            
            # Quick normalization
            if np.max(np.abs(audio_data)) > 0:
                audio_data = audio_data / np.max(np.abs(audio_data))
            
            # Trim silence from start and end
            audio_data, _ = librosa.effects.trim(audio_data, top_db=20)
            
            return audio_data, target_sr
            
        except Exception as e:
            raise Exception(f"Audio preprocessing failed: {e}")
    
    def transcribe_fast(self, audio_input, language_hint=""):
        """Fast transcription with minimal overhead"""
        
        # Lazy load model
        if not self.load_model_lazy():
            return "❌ Model failed to load. Please check your setup."
        
        try:
            # Quick audio processing
            audio, sr = self.preprocess_audio_minimal(audio_input)
            
            # Minimal system prompt for speed
            system_content = "Transcribe audio to text."
            if language_hint and language_hint != "auto":
                system_content += f" Language: {language_hint}."
            
            turns = [
                {'role': 'system', 'content': system_content},
                {'role': 'user', 'content': '<|audio|>'}
            ]
            
            # Fast inference settings
            with torch.inference_mode():  # More efficient than no_grad
                result = self.pipe(
                    {
                        'audio': audio, 
                        'turns': turns, 
                        'sampling_rate': sr
                    },
                    max_new_tokens=128,    # Reduced further
                    do_sample=False,       # Deterministic
                    num_beams=1,          # No beam search
                    early_stopping=True,   # Stop as soon as possible
                    pad_token_id=self.pipe.tokenizer.eos_token_id if hasattr(self.pipe, 'tokenizer') else None
                )
            
            # Immediate cleanup
            del audio
            gc.collect()
            
            # Extract result
            if isinstance(result, list) and len(result) > 0:
                text = result[0].get('generated_text', '').strip()
            elif isinstance(result, dict):
                text = result.get('generated_text', '').strip()
            else:
                text = str(result).strip()
            
            # Clean up the output (remove system prompts if they appear)
            if "Transcribe audio to text" in text:
                text = text.replace("Transcribe audio to text", "").strip()
            if text.startswith("Language:"):
                text = text.split(".", 1)[-1].strip() if "." in text else text
                
            return text if text else "No speech detected"
            
        except Exception as e:
            return f"❌ Transcription error: {str(e)}"

# Initialize ASR system
print("Initializing Ultra-Light Shuka ASR...")
asr_system = UltraLightShukaASR()

def process_audio(audio, language):
    """Main processing function"""
    if audio is None:
        return "Please upload or record an audio file."
    
    return asr_system.transcribe_fast(audio, language)

# Simple language options
LANGUAGES = [
    ("Auto", "auto"),
    ("English", "english"),
    ("Hindi", "hindi"),
    ("Bengali", "bengali"),
    ("Tamil", "tamil"),
    ("Telugu", "telugu"),
    ("Gujarati", "gujarati"),
    ("Kannada", "kannada"),
    ("Malayalam", "malayalam"),
    ("Marathi", "marathi"),
    ("Punjabi", "punjabi"),
    ("Oriya", "oriya")
]

# Ultra-minimal Gradio interface
css = """
.gradio-container {
    max-width: 800px !important;
}
.output-text textarea {
    font-size: 16px !important;
}
"""

with gr.Blocks(css=css, title="Fast Shuka ASR") as demo:
    
    gr.HTML("""
    <div style='text-align: center; margin-bottom: 20px;'>
        <h1>🚀 Ultra-Fast Shuka v1 ASR</h1>
        <p>Optimized for speed • Multilingual • 15-second max clips</p>
    </div>
    """)
    
    with gr.Row():
        with gr.Column(scale=1):
            audio_input = gr.Audio(
                label="🎙️ Audio Input",
                type="filepath",
                format="wav",
                elem_id="audio-input"
            )
            
            language_select = gr.Dropdown(
                choices=LANGUAGES,
                value="auto",
                label="🌍 Language Hint",
                info="Optional - helps with accuracy"
            )
            
        with gr.Column(scale=2):
            output_box = gr.Textbox(
                label="📝 Transcription",
                placeholder="Upload audio to see transcription here...",
                lines=8,
                elem_classes=["output-text"]
            )
            
            gr.Button("🔄 Clear", size="sm").click(
                lambda: ("", None), 
                outputs=[output_box, audio_input]
            )
    
    # Auto-transcribe on upload
    audio_input.change(
        fn=process_audio,
        inputs=[audio_input, language_select],
        outputs=output_box,
        show_progress=True
    )
    
    # Also trigger on language change
    language_select.change(
        fn=process_audio,
        inputs=[audio_input, language_select],
        outputs=output_box,
        show_progress=True
    )
    
    gr.HTML("""
    <div style='margin-top: 20px; padding: 15px; background: #f0f0f0; border-radius: 10px;'>
        <h4>⚡ Speed Optimizations Active:</h4>
        <ul style='margin: 10px 0;'>
            <li>✅ Auto audio trimming (15s max)</li>
            <li>✅ CPU-optimized inference</li>
            <li>✅ Minimal token generation</li>
            <li>✅ Memory cleanup after each request</li>
        </ul>
        <p><strong>Tip:</strong> For fastest results, use short, clear audio clips in WAV format.</p>
    </div>
    """)

if __name__ == "__main__":
    demo.queue(max_size=3)  # Limit concurrent requests
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=False,
        show_error=True,
        quiet=False
    )