src/preprocess.py

#this version was created for hugging face. if issues persist revert to preprocess_local


import librosa
import numpy as np
import os

# --- CONSTANTS ---
SAMPLE_RATE = 22050  
DURATION = 1.0       # Slice length (Seconds)
SAMPLES_PER_SLICE = int(SAMPLE_RATE * DURATION)
N_MELS = 128         # Frequency resolution

def audio_to_spectrograms(file_path):
    """
    Converts audio into a batch of 1-second spectrogram slices.
    Returns shape: (Num_Slices, 128, 44, 1)
    """
    try:
        # 1. Safety Check: File Existence
        if not os.path.exists(file_path):
            print(f"❌ Error: File not found at {file_path}")
            return None

        # 2. Load Audio
        # We enforce sr=22050 for consistency
        y, sr = librosa.load(file_path, sr=SAMPLE_RATE)

        # 3. Calculate Slices
        num_slices = len(y) // SAMPLES_PER_SLICE
        
        # Safety: If audio is too short (< 1 second), fail gracefully
        if num_slices < 1: 
            print("❌ Audio too short")
            return None

        # 4. Create Spectrograms
        spectrograms = []
        for i in range(num_slices):
            # Extract 1-second chunk
            y_slice = y[i*SAMPLES_PER_SLICE : (i+1)*SAMPLES_PER_SLICE]
            
            # Generate Mel Spectrogram
            spec = librosa.feature.melspectrogram(y=y_slice, sr=sr, n_mels=N_MELS)
            
            # Convert to Decibels (Log Scale) and Normalize (0-1)
            log_spec = librosa.power_to_db(spec, ref=np.max)
            norm_spec = np.clip((log_spec + 80) / 80, 0, 1)
            
            # Add Channel Dimension (Required for CNNs)
            spectrograms.append(norm_spec[..., np.newaxis])
            
        return np.array(spectrograms)

    except Exception as e:
        print(f"❌ Preprocessing Error: {e}")
        return None