app.py

import gradio as gr
import os
import numpy as np
import librosa
from resemblyzer import VoiceEncoder
import soundfile as sf

# Initialize the encoder
encoder = VoiceEncoder()
reference_embeddings = []

def load_audio(file, target_sr=16000):
    """Loads an audio file and resamples it to the target sample rate."""
    audio, sr = librosa.load(file, sr=target_sr)
    return audio

def extract_embeddings(encoder, audio_files):
    """Extracts voice embeddings from uploaded audio files."""
    embeddings = []
    for file_path in audio_files:
        audio = load_audio(file_path)  # Load and preprocess the audio file
        embedding = encoder.embed_utterance(audio)  # Extract voice embedding
        embeddings.append(embedding)
    return np.array(embeddings)

def compute_similarity(embedding, reference_embeddings):
    """Computes similarity between test and reference embeddings."""
    similarities = np.dot(reference_embeddings, embedding) / (
        np.linalg.norm(reference_embeddings, axis=1) * np.linalg.norm(embedding)
    )
    return np.mean(similarities)

def train_voice_samples(files):
    global reference_embeddings
    if len(files) > 50:
        return "Please upload up to 50 files only."

    reference_embeddings = extract_embeddings(encoder, [file.name for file in files])
    return f"Extracted embeddings from {len(files)} voice samples. Ready for testing!"

def test_voice(file):
    try:
        if reference_embeddings is None or len(reference_embeddings) == 0:
            return "No reference voice samples found. Please upload training samples first."

        # Debugging: Check if file is received
        print(f"Received test file: {file.name}")

        # Load test audio properly
        test_audio, sr = librosa.load(file.name, sr=16000)

        # Debugging: Check audio shape
        print(f"Loaded test audio, shape: {test_audio.shape}, Sample rate: {sr}")

        # Extract embedding
        test_embedding = encoder.embed_utterance(test_audio)

        # Compute similarity
        similarity_score = compute_similarity(test_embedding, reference_embeddings)

        # Debugging: Check similarity score
        print(f"Computed similarity score: {similarity_score}")

        # Generate result message
        result = f"Similarity Score: {similarity_score:.2f}\n"
        if similarity_score > 0.8:
            result += "The voice matches closely with the training samples!\n"
            result += "You can access the demo content via the following link:\n"
            result += "[🔗 Access Demo](https://example.com/demo)"
        else:
            result += "The voice does not match the training samples."
        return result

    except Exception as e:
        return f"Error: {str(e)}"

with gr.Blocks() as app:
    gr.Markdown("## Voice Recognition with Similarity Testing")
    gr.Markdown("**Instruction:** Upload a single file of more than 1-minute duration or multiple files totaling more than 1 minute.")
    gr.Markdown("[🔗 Link to Eleven Labs](https://elevenlabs.io/app/speech-synthesis/text-to-speech)")
    gr.Markdown("**Access Eleven Labs to test the model on multiple voices**")
    
    with gr.Row():
        train_audio = gr.File(label="Upload up to 50 training voice samples", file_types=[".wav"], file_count="multiple")
        train_button = gr.Button("Train Model")
    train_output = gr.Textbox()
    train_button.click(train_voice_samples, inputs=train_audio, outputs=train_output)

    with gr.Row():
        test_audio = gr.File(label="Upload a test voice file", file_types=[".wav"])
        test_button = gr.Button("Test Voice")
    test_output = gr.Textbox()
    test_button.click(test_voice, inputs=test_audio, outputs=test_output)

app.launch(share=True)