import soundfile as sf import torch from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor import gradio as gr import sox import subprocess import openai # Set your OpenAI API key api_key = "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" # spell_checker = GoogleSpellChecker(lang="fa") def read_file_and_process(wav_file): filename = wav_file.split('.')[0] filename_16k = filename + "16k.wav" resampler(wav_file, filename_16k) speech, _ = sf.read(filename_16k) inputs = processor(speech, sampling_rate=16_000, return_tensors="pt", padding=True) return inputs def resampler(input_file_path, output_file_path): command = ( f"ffmpeg -hide_banner -loglevel panic -i {input_file_path} -ar 16000 -ac 1 -bits_per_raw_sample 16 -vn " f"{output_file_path}" ) subprocess.call(command, shell=True) def parse_transcription(logits): predicted_ids = torch.argmax(logits, dim=-1) transcription = processor.decode(predicted_ids[0], skip_special_tokens=True) del(logits) return transcription # def corrector(sentence): # check_spell = spell_checker.check(sentence) # if check_spell[1] is None: # return sentence # else: # return check_spell[1] def correct_text_with_gpt(text): openai.api_key = api_key response = openai.Completion.create( engine="text-davinci-003", prompt=f"Please correct the following text: '{text}'\n\nCorrected text:", max_tokens=1000, temperature=0.5, # Temperature controls the randomness of the model's output. A higher value like 1.0 makes the output more random, while a lower value like 0.2 makes it more deterministic and focused. top_p=1.0, # This parameter controls the diversity of the output. It sets a threshold for the cumulative probability of words to keep. Smaller values like 0.2 will result in more focused responses, while larger values like 0.8 will allow for more diversity. frequency_penalty=0.2, # encourages the use of less common words presence_penalty=0.5, # discourages the use of common words. ) return response.choices[0].text.strip() def parse(wav_file): input_values = read_file_and_process(wav_file) with torch.no_grad(): logits = model(**input_values).logits return correct_text_with_gpt(parse_transcription(logits)) # def parse(wav_file): # check_spell = '' # input_values = read_file_and_process(wav_file) # with torch.no_grad(): # logits = model(**input_values).logits # # sentence = parse_transcription(logits) # check_spell = spell_checker.check(parse_transcription(logits)) # # if check_spell[0] is False: # # corrected = check_spell[1] # # else: # # corrected = sentence # return spell_checker.check(parse_transcription(logits))[1] if spell_checker.check(parse_transcription(logits))[0] is False else parse_transcription(logits) model_id = "jonatasgrosman/wav2vec2-large-xlsr-53-persian" processor = Wav2Vec2Processor.from_pretrained(model_id) model = Wav2Vec2ForCTC.from_pretrained(model_id) input_ = gr.Audio(source="microphone", type="filepath", label="لطفا دکمه ضبط صدا را بزنید و شروع به صحبت کنید و بعذ از اتمام صحبت دوباره دکمه ضبط را فشار دهید.", show_download_button=True, show_edit_button=True, ) txtbox = gr.Textbox( label="متن گفتار شما: ", lines=5, text_align="right", show_label=True, show_copy_button=True, ) title = "Speech-to-Text (persian)" description = "، توجه داشته باشید که هرچه گفتار شما شمرده تر باشد خروجی با کیفیت تری دارید.روی دکمه ضبط صدا کلیک کنید و سپس دسترسی مرورگر خود را به میکروفون دستگاه بدهید، سپس شروع به صحبت کنید و برای اتمام ضبط دوباره روی دکمه کلیک کنید" article = "

Large-Scale Self- and Semi-Supervised Learning for Speech Translation

" demo = gr.Interface(fn=parse, inputs = input_, outputs=txtbox, title=title, description=description, article = article, streaming=True, interactive=True, analytics_enabled=False, show_tips=False, enable_queue=True) demo.launch(share=True)