app.py

import sys
import os
import time
import torch
import gradio as gr

# --- 1. PATH SETUP ---
current_dir = os.path.dirname(os.path.abspath(__file__))
src_path = os.path.join(current_dir, "src")
if src_path not in sys.path:
    sys.path.append(src_path)

# --- 2. Imports ---
try:
    from kanade_tokenizer.model import KanadeModel
    from kanade_tokenizer.util import load_vocoder, vocode, load_audio
except ImportError as e:
    print(f"❌ IMPORT ERROR: {e}")
    raise e

# --- Configuration ---
KANADE_REPO = "frothywater/kanade-25hz-clean"
KANADE_VOCODER = "hift"
DEVICE = "cpu"
SAMPLE_RATE = 24000
MAX_AUDIO_SECONDS = 30  # Limit audio to 30 seconds

print(f"🚀 Initializing on {DEVICE}...")

# --- 3. Load Models ---
print(f"📥 Loading Kanade...")
kanade_model = KanadeModel.from_pretrained(repo_id=KANADE_REPO).to(DEVICE).eval()

print(f"🔊 Loading HiFT Vocoder...")
kanade_vocoder = load_vocoder(name=KANADE_VOCODER).to(DEVICE).eval()

print("✅ Models Loaded.")

# --- Core Inference ---
def run_inference(source_wav, ref_wav):
    """Run voice conversion inference on CPU"""
    with torch.inference_mode():
        mel_output = kanade_model.voice_conversion(source_wav, ref_wav)
        generated_wav = vocode(kanade_vocoder, mel_output.unsqueeze(0))
    return generated_wav

# --- Main Handler ---
def voice_conversion(source_path, reference_path):
    if not source_path or not reference_path:
        return None, "⚠️ Please provide both source and reference audio."
    
    try:
        # Load audio
        source_wav = load_audio(source_path, sample_rate=SAMPLE_RATE).to(DEVICE)
        ref_wav = load_audio(reference_path, sample_rate=SAMPLE_RATE).to(DEVICE)
        
        # Check duration (30 second limit)
        max_samples = MAX_AUDIO_SECONDS * SAMPLE_RATE
        
        if source_wav.shape[-1] > max_samples:
            source_wav = source_wav[..., :max_samples]
        
        if ref_wav.shape[-1] > max_samples:
            ref_wav = ref_wav[..., :max_samples]
        
        # Run inference
        start = time.time()
        final_wav = run_inference(source_wav, ref_wav)
        proc_time = time.time() - start
        
        output_np = final_wav.squeeze().cpu().float().numpy()
        output_duration = len(output_np) / SAMPLE_RATE
        
        # RTF = processing time / audio duration (lower is better, <1 means faster than real-time)
        rtf = proc_time / output_duration if output_duration > 0 else 0
        
        return (SAMPLE_RATE, output_np), f"✅ {proc_time:.2f}s to convert {output_duration:.1f}s of audio | RTF: {rtf:.2f}x"
    
    except Exception as e:
        import traceback
        traceback.print_exc()
        return None, f"❌ Error: {str(e)}"

# --- Gradio Interface ---
with gr.Blocks(title="Kanade Voice Cloning") as demo:
    gr.Markdown("""
    # 🗣️ Kanade Voice Cloning
    **Model:** `frothywater/kanade-25hz-clean`
    
    Convert any audio into a target voice. Upload a source audio (what to say) and a reference audio (whose voice to use).
    
    ⏱️ **Limit:** Audio is trimmed to 30 seconds max.
    """)
    
    with gr.Row():
        with gr.Column():
            source_audio = gr.Audio(label="Source Audio (Content - what to say)", type="filepath")
            reference_audio = gr.Audio(label="Reference Audio (Target Voice - whose voice)", type="filepath")
            convert_btn = gr.Button("🎤 Convert Voice", variant="primary")
        
        with gr.Column():
            output_audio = gr.Audio(label="Result")
            status_text = gr.Textbox(label="Status", interactive=False)
    
    convert_btn.click(
        voice_conversion, 
        inputs=[source_audio, reference_audio], 
        outputs=[output_audio, status_text]
    )
    
    gr.Markdown("""
    ---
    **Tips:**
    - For best results, use clean reference audio (3-10 seconds of clear speech)
    - Source and reference should ideally be similar in speaking pace
    """)

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