Definition
Whisper is a speech recognition model and open-source implementation released by OpenAI. The official README describes it as “a general-purpose speech recognition model” that performs multilingual speech recognition, speech translation, and language identification; the code and model weights are published under the MIT license in the official GitHub repository.
How it works
It is a Transformer sequence-to-sequence model that represents multilingual recognition, speech translation, language identification, and voice activity detection as token sequences to be predicted. The repository offers six model sizes (four of them English-only) plus a turbo variant optimized from large-v3, trading speed against accuracy. Per the README, running pip install -U openai-whisper with ffmpeg installed lets you transcribe locally.
Usage
Its main use is generating text transcripts from audio files, and it can also translate non-English speech into English. Because the code and weights are open, it is widely used for local transcription, subtitle preparation, and automated speech-processing pipelines. Note that the official README states the turbo model is not intended for translation tasks; use the multilingual models instead.
Related and confusable terms
Whisper is a model and a command-line implementation, not an end-user graphical application; many transcription apps and browser plugins call Whisper or its derivatives internally, and their results depend on that integration. Hosted transcription through OpenAI’s platform is a separate usage; the entry point is the Audio API reference.
Episode discussion
In Weekly #007, in the chapter “更大胆的任务：播客转写、Slack 与微信整理” (Chinese transcript), Orange describes a browser plugin that transcribes podcast subtitles locally, and Xiangyang Qiaomu asks which model it uses, mentioning “Qwen 0.6B and Whisper”; Orange replies that the transcription works well. See the chapter.
Frequently asked questions
Can Whisper run locally?
Yes. Per the official README, install it with pip install -U openai-whisper, install ffmpeg, and transcription runs on your own machine; model files are downloaded for the chosen size.
Does Whisper support translation?
Yes — translating non-English speech into English — but the official README states the turbo model is not trained for translation; use the multilingual models from tiny to large instead.
Is there an official Whisper API?
OpenAI’s platform documents audio transcription interfaces in the Audio API reference; that hosted service is a separate usage from the open-source model.