app.py

import os
import io
import csv
import gradio as gr
import numpy as np
import tensorflow as tf
import tensorflow_hub as hub
import tensorflow_io as tfio
import matplotlib.pyplot as plt
from tensorflow import keras
from huggingface_hub import from_pretrained_keras

# Configuration
class_names = [
    "Irish",
    "Midlands",
    "Northern",
    "Scottish",
    "Southern",
    "Welsh",
    "Not a speech",
]

# Download Yamnet model from TF Hub
yamnet_model = hub.load("https://tfhub.dev/google/yamnet/1")

# Download dense model from HF Hub
model = from_pretrained_keras(
    pretrained_model_name_or_path="fbadine/uk_ireland_accent_classification"
)

# Function that reads a wav audio file and resamples it to 16000 Hz
# This function is copied from the tutorial:
# https://www.tensorflow.org/tutorials/audio/transfer_learning_audio
def load_16k_audio_wav(filename):
    # Read file content
    file_content = tf.io.read_file(filename)

    # Decode audio wave
    audio_wav, sample_rate = tf.audio.decode_wav(file_content, desired_channels=1)
    audio_wav = tf.squeeze(audio_wav, axis=-1)
    sample_rate = tf.cast(sample_rate, dtype=tf.int64)

    # Resample to 16k
    audio_wav = tfio.audio.resample(audio_wav, rate_in=sample_rate, rate_out=16000)

    return audio_wav


# Function thatt takes the audio file produced by gr.Audio(source="microphone") and
# returns a tensor applying the following transformations:
# - Resample to 16000 Hz
# - Normalize
# - Reshape to [1, -1]
def mic_to_tensor(recorded_audio_file):
    sample_rate, audio = recorded_audio_file

    audio_wav = tf.constant(audio, dtype=tf.float32)
    if tf.rank(audio_wav) > 1:
        audio_wav = tf.reduce_mean(audio_wav, axis=1)
    audio_wav = tfio.audio.resample(audio_wav, rate_in=sample_rate, rate_out=16000)

    audio_wav = tf.divide(audio_wav, tf.reduce_max(tf.abs(audio_wav)))

    return audio_wav


# Function that takes a tensor and applies the following:
# - Pass it through Yamnet model to get the embeddings which are the input of the dense model
# - Pass the embeddings through the dense model to get the predictions
def tensor_to_predictions(audio_tensor):
    # Get audio embeddings & scores.
    scores, embeddings, mel_spectrogram = yamnet_model(audio_tensor)

    # Predict the output of the accent recognition model with embeddings as input
    predictions = model.predict(embeddings)

    return predictions, mel_spectrogram


# Function tha is called when the user clicks "Predict" button. It does the following:
# - Calls tensor_to_predictions() to get the predictions
# - Generates the top scoring labels
# - Generates the top scoring plot
def predict_accent(recorded_audio_file, uploaded_audio_file):
    # Transform input to tensor
    if recorded_audio_file:
        audio_tensor = mic_to_tensor(recorded_audio_file)
    else:
        audio_tensor = load_16k_audio_wav(uploaded_audio_file)

    # Model Inference
    predictions, mel_spectrogram = tensor_to_predictions(audio_tensor)

    # Get the infered class
    infered_class = class_names[predictions.mean(axis=0).argmax()]

    # Generate Output 1 - Accents
    top_scoring_labels_output = {
        class_names[i]: float(predictions.mean(axis=0)[i])
        for i in range(len(class_names))
    }

    # Generate Output 2
    top_scoring_plot_output = generate_top_scoring_plot(predictions)

    return [top_scoring_labels_output, top_scoring_plot_output]


# Clears all inputs and outputs when the user clicks "Clear" button
def clear_inputs_and_outputs():
    return [None, None, None, None]


# Function that generates the top scoring plot
# This function is copied from the tutorial and adjusted to our needs
# https://keras.io/examples/audio/uk_ireland_accent_recognition/tinyurl.com/4a8xn7at
def generate_top_scoring_plot(predictions):
    # Plot and label the model output scores for the top-scoring classes.
    mean_predictions = np.mean(predictions, axis=0)

    top_class_indices = np.argsort(mean_predictions)[::-1]
    fig = plt.figure(figsize=(10, 2))
    plt.imshow(
        predictions[:, top_class_indices].T,
        aspect="auto",
        interpolation="nearest",
        cmap="gray_r",
    )

    # patch_padding = (PATCH_WINDOW_SECONDS / 2) / PATCH_HOP_SECONDS
    # values from the model documentation
    patch_padding = (0.025 / 2) / 0.01
    plt.xlim([-patch_padding - 0.5, predictions.shape[0] + patch_padding - 0.5])
    # Label the top_N classes.
    yticks = range(0, len(class_names), 1)
    plt.yticks(yticks, [class_names[top_class_indices[x]] for x in yticks])
    _ = plt.ylim(-0.5 + np.array([len(class_names), 0]))

    return fig


# Main function
if __name__ == "__main__":
    demo = gr.Blocks()

    with demo:
        gr.Markdown(
            """
            <center><h1>English speaker accent recognition using Transfer Learning</h1></center> \
            This space is a demo of an English (precisely UK & Ireland) accent classification model using Keras.<br> \
    In this space, you can record your voice or upload a wav file and the model will predict the English accent spoken in the audio<br><br>
            """
        )
        with gr.Row():
            ## Input
            with gr.Column():
                mic_input = gr.Audio(source="microphone", label="Record your own voice")
                upl_input = gr.Audio(
                    source="upload", type="filepath", label="Upload a wav file"
                )

                with gr.Row():
                    clr_btn = gr.Button(value="Clear", variant="secondary")
                    prd_btn = gr.Button(value="Predict")

            with gr.Column():
                lbl_output = gr.Label(label="Top Predictions")
                with gr.Group():
                    gr.Markdown("<center>Prediction per time slot</center>")
                    plt_output = gr.Plot(
                        label="Prediction per time slot", show_label=False
                    )

        clr_btn.click(
            fn=clear_inputs_and_outputs,
            inputs=[],
            outputs=[mic_input, upl_input, lbl_output, plt_output],
        )
        prd_btn.click(
            fn=predict_accent,
            inputs=[mic_input, upl_input],
            outputs=[lbl_output, plt_output],
        )

    demo.launch(debug=True, share=True)