audio_backend.py

import tensorflow as tf
import numpy as np
from audio_utils import audio_to_spectrogram

# ======================================================
# MODEL LOAD
# ======================================================
MODEL_PATH = "models/audio_vit_savedmodel"

model = tf.saved_model.load(MODEL_PATH)
infer = model.signatures["serving_default"]


# ======================================================
# PREDICTION FUNCTION (UI-SAFE)
# ======================================================
def predict_audio(audio_file_path):
    """
    Returns:
    - label (real / fake)
    - confidence (%)
    - spectrogram image
    - error message (None if OK)
    """

    try:
        # Convert audio → spectrogram
        spec_img = audio_to_spectrogram(audio_file_path)

        x = spec_img.astype("float32") / 255.0
        x = np.expand_dims(x, axis=0)

        preds = infer(tf.constant(x))
        prob = list(preds.values())[0].numpy()[0][0]

        label = "fake" if prob >= 0.5 else "real"
        confidence = round(prob * 100, 2)

        return label, confidence, spec_img, None

    except ValueError as ve:
        # Expected errors (short audio, invalid input)
        return None, None, None, str(ve)

    except Exception:
        # Unexpected errors (decoding/model issues)
        return None, None, None, (
            "Unable to process the audio file. "
            "Please upload a clear audio clip in WAV, MP3, FLAC, or M4A format."
        )