app.py

import gradio as gr
import torch
from transformers import WhisperForConditionalGeneration, WhisperProcessor
import librosa
import numpy as np

# Load model and processor
def load_model():
    try:
        model = WhisperForConditionalGeneration.from_pretrained(
            "therealbee/whisper-small-ha-bible-tts",
            ignore_mismatched_sizes=True
        )
        processor = WhisperProcessor.from_pretrained("therealbee/whisper-small-ha-bible-tts")
        return model, processor
    except Exception as e:
        print(f"Error loading model: {e}")
        return None, None

# Load model once
model, processor = load_model()

def transcribe_audio(audio_input):
    """Transcribe audio file to Hausa text"""
    if model is None or processor is None:
        return "❌ Model failed to load. Please try again later."
    
    if audio_input is None:
        return "❌ Please provide an audio file or record audio."
    
    try:
        # Handle different audio input types
        if isinstance(audio_input, tuple):
            # Recorded audio: (sample_rate, audio_data)
            sample_rate, audio_data = audio_input
            audio = audio_data.astype(np.float32)
            
            # Normalize if needed
            if audio.max() > 1.0:
                audio = audio / np.max(np.abs(audio))
            
            # Resample if needed
            if sample_rate != 16000:
                audio = librosa.resample(audio, orig_sr=sample_rate, target_sr=16000)
        else:
            # Uploaded file: filepath
            audio, sample_rate = librosa.load(audio_input, sr=16000)
        
        # Ensure audio is not empty
        if len(audio) < 1000:  # Less than ~0.06 seconds
            return "❌ Audio too short. Please record at least 1 second of speech."
        
        # Prepare inputs
        inputs = processor(
            audio, 
            sampling_rate=16000, 
            return_tensors="pt", 
            language="ha"
        )
        
        # Generate transcription
        with torch.no_grad():
            outputs = model.generate(inputs.input_features, task="transcribe")
        
        # Decode transcription
        transcription = processor.batch_decode(outputs, skip_special_tokens=True)[0]
        
        if not transcription.strip():
            return "❌ No speech detected. Please speak more clearly or check your audio."
        
        return f"📝 Hausa Transcription:\n\n{transcription}"
        
    except Exception as e:
        return f"❌ Transcription failed: {str(e)}"

# Create Gradio interface
demo = gr.Interface(
    fn=transcribe_audio,
    inputs=gr.Audio(
        sources=["microphone", "upload"],
        type="numpy",
        label="🎙️ Record or Upload Hausa Audio"
    ),
    outputs=gr.Textbox(
        label="📋 Transcription Result",
        lines=5,
        placeholder="Your Hausa transcription will appear here..."
    ),
    title="🎙️ Hausa Speech Transcription",
    description="""
    Upload an audio file or record directly to get Hausa transcription.
    
    **Supported formats:** WAV, MP3, OGG, M4A, FLAC
    
    **Tips:**
    - Speak clearly in Hausa
    - Keep recordings under 30 seconds for best results
    - Use good quality audio
    """,
    examples=[],
    theme=gr.themes.Soft(),
    allow_flagging="never"
)

if __name__ == "__main__":
    demo.launch()