app.py

import gradio as gr
import librosa
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
from matplotlib.backends.backend_agg import FigureCanvasAgg
from keras.models import load_model

# Load the Keras model
model = load_model('test.h5')

# Define classes
classes = ['Class 1', 'Class 2', 'Class 3', 'Class 4', 'Class 5']

def classify_syllable(voice):
    # Convert voice to spectrogram image
    img = voice_to_image(voice)
    # Perform prediction
    prediction = model.predict(img[np.newaxis, ..., np.newaxis])[0]
    # Get class label
    predicted_class = classes[np.argmax(prediction)]
    # Plot spectrogram
    plot = plot_spectrogram(img)
    # Return prediction and spectrogram plot
    return predicted_class, plot

def voice_to_image(voice):
    audio_data, sample_rate = librosa.load(voice, sr=None)
    spectrogram = librosa.feature.melspectrogram(y=audio_data, sr=sample_rate)
    spectrogram_db = librosa.power_to_db(spectrogram, ref=np.max)
    fig, ax = plt.subplots()
    ax.imshow(spectrogram_db, aspect='auto', cmap='viridis')
    ax.axis('off')
    fig.tight_layout(pad=0)
    return fig

def plot_spectrogram(img):
    canvas = FigureCanvasAgg(img)
    canvas.draw()
    image_data = np.frombuffer(canvas.tostring_rgb(), dtype='uint8')
    image_data = image_data.reshape(img.get_width_height()[::-1] + (3,))
    return image_data

# Define Gradio interface
voice_input = gr.inputs.Audio(source="microphone", type="file", label="Record Voice")
submit_button = gr.inputs.Button(text="Submit")
output_label = gr.outputs.Label(label="Predicted Class")
output_image = gr.outputs.Image(type="numpy", label="Spectrogram")

gr.Interface(
    fn=classify_syllable,
    inputs=voice_input,
    outputs=[output_label, output_image],
    title="Malay Language Syllable Classification",
    description="Record your voice and classify the Malay language syllable into one of five classes.",
    theme="compact",
    allow_flagging=False,
    allow_screenshot=False,
    allow_browser_cache=False
).launch()