I am pleased to announce that OpenAI has released another product from its remarkable workshop...

I am pleased to announce that OpenAI has released another product from its remarkable workshop. It is called 'Whisper'. Whisper is a universal end-to-end, weakly supervised family of ASR (Automatic Speech Recognition) models based on transformers. In simple terms, it converts spoken words into text. And it does so in a way we refer to as "General-purpose", meaning that in addition to speech recognition, it can also perform tasks such as voice detection, language identification, transcription, and machine translation.

A whole family of models is being released, varying in size. From the smallest to the largest in terms of parameters: Tiny (39M), Base (74M), Small (244M), Medium (769M), and Large (1.55B). The great news is that it is being released as open source! Interested parties can easily try out the online demo and see how the model works in Czech (or those more curious or skilled can download it directly from GitHub). And a little cherry on top: the models are trained on 77 years of spoken discourse collected from the internet, which I believe is the largest similar dataset of its kind.

Sources:

• Demo on Hugging Face: https://huggingface.co/spaces/openai/whisper
• OpenAI blog: https://openai.com/blog/whisper/
• Paper: https://cdn.openai.com/papers/whisper.pdf
• GitHub: https://github.com/openai/whisper
• Colab: https://colab.research.google.com/github/openai/whisper/blob/master/notebooks/LibriSpeech.ipynb
• Medium: https://towardsdatascience.com/openai-whisper-holds-the-key-to-gpt-4-a7f922a7dad9

Originally published on Facebook — link to post