import gradio as gr from transformers import pipeline import numpy as np import pandas as pd import re from pydub import AudioSegment from pydub.generators import Sine import io from scipy.signal import resample MODEL_NAME = "openai/whisper-tiny" BATCH_SIZE = 8 # device = 0 if torch.cuda.is_available() else "cpu" pipe = pipeline( task="automatic-speech-recognition", model=MODEL_NAME, chunk_length_s=30, # device=device, ) arabic_bad_Words = pd.read_csv("arabic_bad_words_dataset.csv") english_bad_Words = pd.read_csv("english_bad_words_dataset.csv") def clean_english_word(word): # Use regex to remove special characters, punctuation, and spaces around words cleaned_text = re.sub(r'^[\s\W_]+|[\s\W_]+$', '', word) return cleaned_text def clean_arabic_word(word): # Define a regex pattern to match any non-Arabic letter character pattern = r'[^\u0600-\u06FF]' # Replace any character matching the pattern with an empty string cleaned_word = re.sub(pattern, '', word) return cleaned_word def classifier(word_list_with_timestamp, language): foul_words = [] negative_timestamps = [] if language == "English": list_to_search = set(english_bad_Words["words"]) for item in word_list_with_timestamp: word = clean_english_word(item['text']) if word in list_to_search: foul_words.append(word) negative_timestamps.append(item['timestamp']) else: list_to_search = list(arabic_bad_Words["words"]) for item in word_list_with_timestamp: word = clean_arabic_word(item['text']) for word_in_list in list_to_search: if word_in_list == word: foul_words.append(word) negative_timestamps.append(item['timestamp']) break return [foul_words, negative_timestamps] def generate_bleep(duration_ms, frequency=1000): sine_wave = Sine(frequency) bleep = sine_wave.to_audio_segment(duration=duration_ms) return bleep def mute_audio_range(audio_filepath, ranges, bleep_frequency=800): audio = AudioSegment.from_file(audio_filepath) for range in ranges: start_time = range[0] - 0.1 end_time = range[-1] + 0.1 start_ms = start_time * 1000 # pydub works with milliseconds end_ms = end_time * 1000 duration_ms = end_ms - start_ms # Generate the bleep sound bleep_sound = generate_bleep(duration_ms, bleep_frequency) # Combine the original audio with the bleep sound audio = audio[:start_ms] + bleep_sound + audio[end_ms:] return audio def format_output_to_list(data): formatted_list = "\n".join([f"{item['timestamp'][0]}s - {item['timestamp'][1]}s \t : {item['text']}" for item in data]) return formatted_list def transcribe(input_audio, audio_language, task, timestamp_type): if input_audio is None: raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.") if timestamp_type == "sentence": timestamp_type = True else: timestamp_type = "word" output = pipe(input_audio, batch_size=BATCH_SIZE, return_timestamps=timestamp_type, generate_kwargs={"task": task}) text = output['text'] timestamps = format_output_to_list(output['chunks']) foul_words, negative_timestamps = classifier(output['chunks'], audio_language) foul_words = ", ".join(foul_words) audio_output = mute_audio_range(input_audio, negative_timestamps) # Save the output audio to a BytesIO object output_buffer = io.BytesIO() audio_output.export(output_buffer, format="wav") output_buffer.seek(0) # Read the audio data from the BytesIO buffer sample_rate = audio_output.frame_rate audio_data = np.frombuffer(output_buffer.read(), dtype=np.int16) return [text, timestamps, foul_words, (sample_rate, audio_data)] examples = [ ["arabic_english_audios/audios/arabic_audio_1.wav", 'Arabic', 'transcribe', 'word'], ["arabic_english_audios/audios/arabic_audio_2.wav", 'Arabic', 'transcribe', 'word'], ["arabic_english_audios/audios/arabic_audio_3.wav", 'Arabic', 'transcribe', 'word'], ["arabic_english_audios/audios/arabic_audio_4.wav", 'Arabic', 'transcribe', 'word'], ["arabic_english_audios/audios/arabic_hate_audio_1.mp3", 'Arabic', 'transcribe', 'word'], ["arabic_english_audios/audios/arabic_hate_audio_2.mp3", 'Arabic', 'transcribe', 'word'], ["arabic_english_audios/audios/arabic_hate_audio_3.mp3", 'Arabic', 'transcribe', 'word'], ["arabic_english_audios/audios/english_audio_1.wav", 'English', 'transcribe', 'word'], ["arabic_english_audios/audios/english_audio_2.mp3", 'English', 'transcribe', 'word'], ["arabic_english_audios/audios/english_audio_3.mp3", 'English', 'transcribe', 'word'], ["arabic_english_audios/audios/english_audio_4.mp3", 'English', 'transcribe', 'word'], ["arabic_english_audios/audios/english_audio_5.mp3", 'English', 'transcribe', 'word'], ["arabic_english_audios/audios/english_audio_6.wav", 'English', 'transcribe', 'word'] ] with gr.Blocks(theme=gr.themes.Default()) as demo: gr.HTML("

Transcribing Audio with Timestamps using whisper-large-v3

") # gr.Markdown("") with gr.Row(): with gr.Column(): audio_input = gr.Audio(sources=["upload", 'microphone'], type="filepath", label="Audio file") audio_language = gr.Radio(["Arabic", "English"], label="Audio Language") task = gr.Radio(["transcribe", "translate"], label="Task") timestamp_type = gr.Radio(["sentence", "word"], label="Timestamp Type") with gr.Row(): clear_button = gr.ClearButton(value="Clear") submit_button = gr.Button("Submit", variant="primary", ) with gr.Column(): transcript_output = gr.Text(label="Transcript") timestamp_output = gr.Text(label="Timestamps") foul_words = gr.Text(label="Foul Words") output_audio = gr.Audio(label="Output Audio", type="numpy") examples = gr.Examples(examples, inputs=[audio_input, audio_language, task, timestamp_type], outputs=[transcript_output, timestamp_output, foul_words, output_audio], fn=transcribe, examples_per_page=20) submit_button.click(fn=transcribe, inputs=[audio_input, audio_language, task, timestamp_type], outputs=[transcript_output, timestamp_output, foul_words, output_audio]) clear_button.add([audio_input, audio_language, task, timestamp_type, transcript_output, timestamp_output, foul_words, output_audio]) if __name__ == "__main__": demo.launch()