OpenAI today releases Whisper, an open-source neural network for automatic speech recognition, along with models and inference code. The model was trained on 680,000 hours of multilingual and multitask audio data collected from the internet, using weak supervision — meaning it simply learned to predict transcripts without explicit alignment or language labels. According to the accompanying paper, Whisper approaches human-level robustness and accuracy on English speech recognition and can automatically transcribe and translate languages including Spanish, Italian, and Japanese, all in a zero-shot setting without fine-tuning.

The Verge reported that the model installs easily on a local machine and can transcribe audio privately without sending files to the cloud. Journalist Peter Sterne and developer Christina Warren announced they are building Stage Whisper, a free transcription app based on Whisper, aiming to give journalists a secure alternative to cloud services. Sterne told The Verge the model produced the best transcription he had used aside from human transcribers.

The release marks a departure from OpenAI’s recent cautious approach to model access: the company has limited access to GPT-3 and DALL-E, citing a desire to learn more about real-world use and continue to iterate on its safety systems. Whisper’s release may signal a different strategy for speech technology, but the blog post framing remains focused on research.