app.py

import gradio as gr
import numpy as np
import librosa
import hashlib
import json
import os

from scipy.io.wavfile import write as wav_write
from check import transcibe, estimate_audio_quality

def visible():
    return gr.update(visible=True),gr.update(visible=True)

def check(audio:gr.Audio):

    sr, audio = audio
    audio = audio.astype(np.float32)
    if len(audio.shape) > 2 and audio.shape[1] > 1:
        audio = np.mean(audio, axis=1)

    quality, quality_score, features = estimate_audio_quality(audio, sr)
    audio_resampled = librosa.resample(audio, orig_sr=sr, target_sr=16000)
    transcription = transcibe(audio_resampled, 16000)

    check_result = f"""
    ### Results

    Qualtiy (0-1):{quality_score}   - {quality}

    Feautes:
    """

    for key, value in features.items():
        check_result += f"""
    - {key}:   {round(value,3)}
    """

    return transcription, check_result

def generate_sample_id(audio, transcription):
    # Combine the audio and transcription into a single string
    combined = f"{audio.tostring()}{transcription}"
    
    # Generate a hash of the combined string
    sample_id = hashlib.sha256(combined.encode('utf-8')).hexdigest()
    
    return sample_id[:10]  # Return the first 10 characters of the hash as the ID


def save_sample(audio, transcription, check_result):
    sample_id = generate_sample_id(audio, transcription)

    # Create a directory to save the sample
    directory = f"{sample_id}_data"
    if not os.path.exists(directory):
        os.makedirs(directory)
    
    # Save the audio as a .wav file
    audio_filename = os.path.join(directory, f"{sample_id}.wav")
    wav_write(audio_filename, 16000, audio)  # Assuming a sample rate of 16kHz
    
    # Save the transcription as a .txt file
    transcription_filename = os.path.join(directory, f"{sample_id}_transcription.txt")
    with open(transcription_filename, 'w') as f:
        f.write(transcription)
    
    # Save the check_result as a JSON file
    check_result_filename = os.path.join(directory, f"{sample_id}_features.json")
    with open(check_result_filename, 'w') as f:
        json.dump(check_result, f)

with gr.Blocks() as app:

    gr.Markdown("# Open ASR Dataset")
    gr.Markdown("   ")
    gr.Markdown("This App is an effort to collectively crowdsource a new Dataset for the ASR community")
    # gr.Markdown("You can Access the Dataset (here)[LINK]")
    gr.Markdown("The Dataset will be updated every 100 created samples")
    gr.Markdown("   ")
    gr.Markdown("Create a New Sample")


    new_audio = gr.Audio(
        label = "Add Audio",
        sources=['upload','microphone']
    )
    with gr.Row():

        language = gr.Radio(
            label="Spoken Language",
            choices=["English","French","Spanish","German","Italian", "Chinese", "Japanese", "Arabic", "Russian", "Portuguese"],
            value= "English",
            scale=4
        )

        bckgrnd_noise = gr.Radio(
            label="Background Noise Level",
            choices=["Quiet", "Moderate", "Noisy"],
            value="Moderate",
            scale=2
        )
        
        recording_env = gr.Radio(
            label="Recording Environment",
            choices=["Studio", "Home", "Outdoors", "Office"],
            value="Home",
            scale=2
        )

    check_sample_btn = gr.Button(
        value="Check Sample",
        variant="secondary",
        size="sm"
    )
    
    with gr.Row():

        check_result = gr.Markdown()

        transcription = gr.TextArea(
            label="Transcription",
            visible=False,
            interactive=True
        )

    save_sample_button = gr.Button(
        value="Save Sample",
        variant="primary",
        size="sm",
        scale=1,
        visible=False
    )

    check_sample_btn.click(visible, outputs=[transcription, save_sample_button])
    check_sample_btn.click(check, inputs=[new_audio], outputs=[transcription, check_result])
    
    save_sample_button.click(save_sample, inputs=[new_audio, transcription, check_result])


app.launch()