app.py

from transformers import pipeline
import gradio as gr
import torch
import numpy as np
import librosa
import matplotlib.pyplot as plt
import noisereduce

model_id = "fydhfzh/hubert-classifier-aug-fold-3"
pipe = pipeline("audio-classification", model=model_id)

def get_binary_values():
	binary_values = []

	for i in range(1, 29):
		binary_rep = format(i, '05b')
		for i in range(1, 4):
			binary_harakat = format(i, '02b')
			binary_values.append(binary_rep + binary_harakat)

	return binary_values

binary_values = get_binary_values()

arabic_letters = [
	"أَ", "إِ", "أُ",
	"بَ", "بِ", "بُ",
	"تَ", "تِ", "تُ",
	"ثَ", "ثِ", "ثُ",
	"جَ", "جِ", "جُ",
	"حَ", "حِ", "حُ",
	"خَ", "خِ", "خُ",
	"دَ", "دِ", "دُ",
	"ذَ", "ذِ", "ذُ",
	"رَ", "رِ", "رُ",
	"زَ", "زِ", "زُ",
	"سَ", "سِ", "سُ",
	"شَ", "شِ", "شُ",
	"صَ", "صِ", "صُ",
	"ضَ", "ضِ", "ضُ",
	"طَ", "طِ", "طُ",
	"ظَ", "ظِ", "ظُ",
	"عَ", "عِ", "عُ",
	"غَ", "غِ", "غُ",
	"فَ", "فِ", "فُ",
	"قَ", "قِ", "قُ",
	"كَ", "كِ", "كُ",
	"لَ", "لِ", "لُ",
	"مَ", "مِ", "مُ",
	"نَ", "نِ", "نُ",
	"هَ", "هِ", "هُ",
	"وَ", "وِ", "وُ",
	"يَ", "يِ", "يُ"
]

arabic_representation = dict(zip(binary_values, arabic_letters))
arabic_representation

def split_input(raw_input):
	mse = librosa.feature.rms(y=raw_input, frame_length=2048, hop_length=512) ** 2
	mse_db = librosa.core.power_to_db(mse.squeeze(), ref=np.min, top_db=None)
	mse_db = mse_db[mse_db != 0]
	
	percentile_param = 10
	extra_db_param = 0
	
	threshold = np.percentile(mse_db, percentile_param) + extra_db_param
	print(threshold)

	intervals = librosa.effects.split(y=raw_input, top_db=threshold) # top_db = 60 - threshold
	splitted_input = []
	
	for i, (start, end) in enumerate(intervals):
		# Add overlapping frames both for trail and lead to ensure good split result
		overlap = 2000
		start = start - overlap if start - overlap >= 0 else 0 
		end = end + overlap if end + overlap <= len(raw_input) else len(raw_input)
		split_audio = raw_input[start:end]

		if len(split_audio) < 16000:
			side_len = (16000 - len(split_audio))/2
			pad_width = (int(side_len), int(side_len))
			split_audio = np.pad(split_audio, pad_width=pad_width, mode='constant', constant_values=(0, 0))
		else:
			split_audio = split_audio[0:16000]

		splitted_input.append(split_audio)
		
	return splitted_input

def process_audio(filepath):
	audio, sr = librosa.load(filepath, sr=16000)
	audio = noisereduce.reduce_noise(audio, sr)
	audio = librosa.util.normalize(audio)
	audios = split_input(audio)
	
	return audios
	

def classify_utterances(filepath):
	audios = process_audio(filepath)
	output = [pipe(audio)[0] for audio in audios]
	predictions = [arabic_representation[x['label']] for x in output]  

	return ' '.join(predictions)

demo = gr.Blocks()

mic_classification = gr.Interface(
	fn=classify_utterances,
	inputs=gr.Audio(sources='microphone', type='filepath'),
	outputs=gr.Textbox()
)

file_classification = gr.Interface(
	fn=classify_utterances,
	inputs=gr.Audio(sources='upload', type='filepath'),
	outputs=gr.Textbox()
)

with demo:
    gr.TabbedInterface(
        [mic_classification, file_classification],
        ['Classify Microphone', 'Classify Audio File']
    )

demo.launch()