main.py

from fastapi import FastAPI, HTTPException, UploadFile, File
from pydantic import BaseModel
import torch
import librosa
import numpy as np
import os
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
import tempfile
import shutil
from dotenv import load_dotenv
import uvicorn
import scipy.spatial.distance as distance

# Load environment variables
load_dotenv()
HF_TOKEN = os.getenv("HF_TOKEN")

app = FastAPI(title="Quran Recitation Comparer API")

class ComparisonResult(BaseModel):
    similarity_score: float
    interpretation: str

# Custom implementation of DTW to replace librosa.sequence.dtw
def custom_dtw(X, Y, metric='euclidean'):
    """
    Custom Dynamic Time Warping implementation.
    
    Args:
        X: First sequence
        Y: Second sequence
        metric: Distance metric ('euclidean' or 'cosine')
        
    Returns:
        D: Cost matrix
        wp: Warping path
    """
    # Get sequence lengths
    n, m = len(X), len(Y)
    
    # Initialize cost matrix
    D = np.zeros((n + 1, m + 1))
    D[0, 1:] = np.inf
    D[1:, 0] = np.inf
    D[0, 0] = 0
    
    # Fill cost matrix
    for i in range(1, n + 1):
        for j in range(1, m + 1):
            if metric == 'euclidean':
                cost = np.sum((X[i-1] - Y[j-1])**2)
            elif metric == 'cosine':
                cost = 1 - np.dot(X[i-1], Y[j-1]) / (np.linalg.norm(X[i-1]) * np.linalg.norm(Y[j-1]))
            D[i, j] = cost + min(D[i-1, j], D[i, j-1], D[i-1, j-1])
    
    # Backtracking
    wp = [(n, m)]
    i, j = n, m
    while i > 0 or j > 0:
        if i == 0:
            j -= 1
        elif j == 0:
            i -= 1
        else:
            min_idx = np.argmin([D[i-1, j-1], D[i-1, j], D[i, j-1]])
            if min_idx == 0:
                i -= 1
                j -= 1
            elif min_idx == 1:
                i -= 1
            else:
                j -= 1
        wp.append((i, j))
    
    wp.reverse()
    return D, wp

class QuranRecitationComparer:
    def __init__(self, model_name="jonatasgrosman/wav2vec2-large-xlsr-53-arabic", token=None):
        """Initialize the Quran recitation comparer with a specific Wav2Vec2 model."""
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        print(f"Using device: {self.device}")

        # Load model and processor once during initialization
        if token:
            print(f"Loading model {model_name} with token...")
            self.processor = Wav2Vec2Processor.from_pretrained(model_name, use_auth_token=token)
            self.model = Wav2Vec2ForCTC.from_pretrained(model_name, use_auth_token=token)
        else:
            print(f"Loading model {model_name} without token...")
            self.processor = Wav2Vec2Processor.from_pretrained(model_name)
            self.model = Wav2Vec2ForCTC.from_pretrained(model_name)

        self.model = self.model.to(self.device)
        self.model.eval()

        # Cache for embeddings to avoid recomputation
        self.embedding_cache = {}
        print("Model loaded successfully!")

    def load_audio(self, file_path, target_sr=16000, trim_silence=True, normalize=True):
        """Load and preprocess an audio file."""
        if not os.path.exists(file_path):
            raise FileNotFoundError(f"Audio file not found: {file_path}")

        print(f"Loading audio: {file_path}")
        y, sr = librosa.load(file_path, sr=target_sr)

        if normalize:
            y = librosa.util.normalize(y)

        if trim_silence:
            # Use librosa.effects.trim which should be available in most versions
            y, _ = librosa.effects.trim(y, top_db=30)

        return y

    def get_deep_embedding(self, audio, sr=16000):
        """Extract frame-wise deep embeddings using the pretrained model."""
        input_values = self.processor(
            audio,
            sampling_rate=sr,
            return_tensors="pt"
        ).input_values.to(self.device)

        with torch.no_grad():
            outputs = self.model(input_values, output_hidden_states=True)

        hidden_states = outputs.hidden_states[-1]
        embedding_seq = hidden_states.squeeze(0).cpu().numpy()

        return embedding_seq

    def compute_dtw_distance(self, features1, features2):
        """Compute the DTW distance between two sequences of features."""
        D, wp = custom_dtw(X=features1, Y=features2, metric='euclidean')
        distance = D[-1, -1]
        normalized_distance = distance / len(wp)
        return normalized_distance

    def interpret_similarity(self, norm_distance):
        """Interpret the normalized distance value."""
        if norm_distance == 0:
            result = "The recitations are identical based on the deep embeddings."
            score = 100
        elif norm_distance < 1:
            result = "The recitations are extremely similar."
            score = 95
        elif norm_distance < 5:
            result = "The recitations are very similar with minor differences."
            score = 80
        elif norm_distance < 10:
            result = "The recitations show moderate similarity."
            score = 60
        elif norm_distance < 20:
            result = "The recitations show some noticeable differences."
            score = 40
        else:
            result = "The recitations are quite different."
            score = max(0, 100 - norm_distance)

        return result, score

    def get_embedding_for_file(self, file_path):
        """Get embedding for a file, using cache if available."""
        if file_path in self.embedding_cache:
            print(f"Using cached embedding for {file_path}")
            return self.embedding_cache[file_path]

        print(f"Computing new embedding for {file_path}")
        audio = self.load_audio(file_path)
        embedding = self.get_deep_embedding(audio)

        # Store in cache for future use
        self.embedding_cache[file_path] = embedding
        print(f"Embedding shape: {embedding.shape}")

        return embedding

    def predict(self, file_path1, file_path2):
        """
        Predict the similarity between two audio files.
        This method can be called repeatedly without reloading the model.

        Args:
            file_path1 (str): Path to first audio file
            file_path2 (str): Path to second audio file

        Returns:
            float: Similarity score
            str: Interpretation of similarity
        """
        print(f"Comparing {file_path1} and {file_path2}")
        # Get embeddings (using cache if available)
        embedding1 = self.get_embedding_for_file(file_path1)
        embedding2 = self.get_embedding_for_file(file_path2)

        # Compute DTW distance
        print("Computing DTW distance...")
        norm_distance = self.compute_dtw_distance(embedding1.T, embedding2.T)
        print(f"Normalized distance: {norm_distance}")

        # Interpret results
        interpretation, similarity_score = self.interpret_similarity(norm_distance)
        print(f"Similarity score: {similarity_score}, Interpretation: {interpretation}")

        return similarity_score, interpretation

    def clear_cache(self):
        """Clear the embedding cache to free memory."""
        self.embedding_cache = {}
        print("Embedding cache cleared")

# Global variable for the comparer instance
comparer = None

@app.on_event("startup")
async def startup_event():
    """Initialize the model when the application starts."""
    global comparer
    print("Initializing model... This may take a moment.")
    try:
        comparer = QuranRecitationComparer(
            model_name="jonatasgrosman/wav2vec2-large-xlsr-53-arabic",
            token=HF_TOKEN
        )
        print("Model initialized and ready for predictions!")
    except Exception as e:
        print(f"Error initializing model: {str(e)}")
        raise

@app.get("/")
async def root():
    """Root endpoint to check if the API is running."""
    return {"message": "Quran Recitation Comparer API is running", "status": "active"}

@app.post("/compare", response_model=ComparisonResult)
async def compare_files(
    file1: UploadFile = File(...),
    file2: UploadFile = File(...)
):
    """
    Compare two audio files and return similarity metrics.
    
    - **file1**: First audio file (MP3, WAV, etc.)
    - **file2**: Second audio file (MP3, WAV, etc.)
    
    Returns similarity score and interpretation.
    """
    if not comparer:
        raise HTTPException(status_code=500, detail="Model not initialized. Please try again later.")
    
    print(f"Received files: {file1.filename} and {file2.filename}")
    temp_dir = tempfile.mkdtemp()
    print(f"Created temporary directory: {temp_dir}")
    
    try:
        # Save uploaded files to temporary directory
        temp_file1 = os.path.join(temp_dir, file1.filename)
        temp_file2 = os.path.join(temp_dir, file2.filename)
        
        with open(temp_file1, "wb") as f:
            content = await file1.read()
            f.write(content)
            
        with open(temp_file2, "wb") as f:
            content = await file2.read()
            f.write(content)
        
        print(f"Files saved to: {temp_file1} and {temp_file2}")
            
        # Compare the files
        similarity_score, interpretation = comparer.predict(temp_file1, temp_file2)
        
        return ComparisonResult(
            similarity_score=similarity_score,
            interpretation=interpretation
        )
    
    except Exception as e:
        print(f"Error processing files: {str(e)}")
        raise HTTPException(status_code=500, detail=f"Error processing files: {str(e)}")
    
    finally:
        # Clean up temporary files
        print(f"Cleaning up temporary directory: {temp_dir}")
        shutil.rmtree(temp_dir, ignore_errors=True)

@app.post("/clear-cache")
async def clear_cache():
    """Clear the embedding cache to free memory."""
    if not comparer:
        raise HTTPException(status_code=500, detail="Model not initialized.")
    
    comparer.clear_cache()
    return {"message": "Embedding cache cleared successfully"}

if __name__ == "__main__":
    uvicorn.run("main:app", host="0.0.0.0", port=7860, log_level="info")