from transformers import pipeline import gradio as gr import whisper wav2vec_en_model = pipeline("automatic-speech-recognition", model="facebook/wav2vec2-base-960h") wav2vec_fr_model = pipeline("automatic-speech-recognition", model="facebook/wav2vec2-large-xlsr-53-french") whisper_model = whisper.load_model("base") def transcribe_audio(language=None, mic=None, file=None): print(language) if mic is not None: audio = mic elif file is not None: audio = file else: return "You must either provide a mic recording or a file" wav2vec_model = load_models(language) transcription = wav2vec_model(audio)["text"] transcription2 = whisper_model.transcribe(audio, language=language)["text"] return transcription, transcription2 def load_models(lang): if lang == 'en': return wav2vec_en_model elif lang == 'fr': return wav2vec_fr_model else: # default english return wav2vec_en_model title = "Speech2text comparison (Wav2vec vs Whisper)" description = """ This Space allows easy comparisons for transcribed texts between Facebook's Wav2vec model and newly released OpenAI's Whisper model.\n (Even if Whisper includes a language detection, here we have decided to select the language to speed up the computation and to focus only on the quality of the transcriptions. The default language is english) """ article = "Check out [the OpenAI Whisper model](https://github.com/openai/whisper) and [the Facebook Wav2vec model](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) that this demo is based off of." examples = [["english_sentence.flac"], ["2022-a-Droite-un-fauteuil-pour-trois-3034044.mp3000.mp3"]] gr.Interface( fn=transcribe_audio, inputs=[ gr.Radio(label="Language", choices=["en", "fr"], value="en"), gr.Audio(source="microphone", type="filepath", optional=True), gr.Audio(source="upload", type="filepath", optional=True), ], outputs=[ gr.Textbox(label="facebook/wav2vec"), gr.Textbox(label="openai/whisper"),], title=title, description=description, article=article, examples=examples ).launch(debug=True)