app/utils.py

import base64
import io
import librosa
import numpy as np
import torch
import torchaudio
import soundfile as sf

def decode_audio(base64_string: str):
    """
    Decodes a base64 string into an in-memory audio file-like object.
    """
    try:
        audio_data = base64.b64decode(base64_string)
        return io.BytesIO(audio_data)
    except Exception as e:
        raise ValueError(f"Invalid Base64 audio data: {str(e)}")

def load_audio(file_obj, target_sr=16000):
    """
    Loads audio from a file object using librosa/torchaudio.
    Returns:
        waveform (torch.Tensor): Audio waveform
        sr (int): Sample rate
    """
    # Load using librosa for robust format handling (MP3, etc)
    y, sr = librosa.load(file_obj, sr=target_sr)
    
    # Noise Reduction (Basic spectral gating) to reduce false positives from background noise
    try:
        import noisereduce as nr
        # Assume noise is estimated from the whole clip (stationary)
        y = nr.reduce_noise(y=y, sr=sr, stationary=True, prop_decrease=0.75)
    except Exception as e:
        print(f"Warning: Noise reduction failed: {e}")

    # Convert to tensor
    waveform = torch.tensor(y).unsqueeze(0) # (1, time)
    return waveform, sr

def extract_heuristic_features(y, sr):
    """
    Extracts simple spectral features for explainability.
    """
    # Spectral Centroid
    cent = librosa.feature.spectral_centroid(y=y, sr=sr)
    mean_cent = np.mean(cent)
    
    # Spectral Rolloff
    rolloff = librosa.feature.spectral_rolloff(y=y, sr=sr)
    mean_rolloff = np.mean(rolloff)
    
    # Zero Crossing Rate
    zcr = librosa.feature.zero_crossing_rate(y)
    mean_zcr = np.mean(zcr)
    
    return {
        "spectral_centroid": float(mean_cent),
        "spectral_rolloff": float(mean_rolloff),
        "zero_crossing_rate": float(mean_zcr)
    }