Description
< #Speech Recognition #Speech to Text #Multilingual #Subtitles #Self-hosted #Open Source
Whisper is OpenAI's open-source general-purpose speech recognition model, and it turns audio into text accurately while being free and able to run on your own machine. Comparable hosted transcription services bill monthly — Otter and its peers run around twenty dollars a month — whereas Whisper keeps working once installed, and the audio never has to leave your computer.
Trained on a large and diverse audio dataset, it recognizes more than ninety languages and is a multitask model besides: alongside transcription it can translate non-English speech straight into English and identify which language is being spoken. It fits meeting notes, podcast transcripts, video subtitles, or adding transcription to your own software at no cost.
Multilingual recognition: over ninety languages are supported, with accuracy varying by language — the project publishes a word-error-rate breakdown per language for reference.
Speech translation: adding `--task translate` turns non-English speech directly into English text, with the multilingual models (small, medium, large) performing best at it.
Six model sizes: from 39M tiny to 1550M large, needing roughly 1GB to 10GB of VRAM across a tenfold speed range, so you can trade speed against accuracy to fit your hardware; the turbo model is an optimized large-v3 that runs about eight times faster with minimal accuracy loss.
English-only variants: tiny, base, small and medium each ship an `.en` model that outperforms the multilingual model of the same size on purely English audio.
Command line ready: `whisper audio.mp3 --model turbo` transcribes in one command, handling flac, mp3, wav and other common formats, with an option to specify the language and to process several files at once.
Python API: `whisper.load_model()` plus `model.transcribe()` is all it takes to embed transcription into your own program or batch pipeline.
Simple installation: `pip install -U openai-whisper` with ffmpeg available on the system, running on Python 3.8 through 3.11.
Runs locally: inference happens on your machine and audio is never uploaded anywhere, which matters for recorded meetings and interview material.
Open source project: the source is hosted on GitHub under the MIT license, so developers can review the implementation, contribute, or customize it for their own needs.
Whisper is OpenAI's open-source general-purpose speech recognition model, and it turns audio into text accurately while being free and able to run on your own machine. Comparable hosted transcription services bill monthly — Otter and its peers run around twenty dollars a month — whereas Whisper keeps working once installed, and the audio never has to leave your computer.
Trained on a large and diverse audio dataset, it recognizes more than ninety languages and is a multitask model besides: alongside transcription it can translate non-English speech straight into English and identify which language is being spoken. It fits meeting notes, podcast transcripts, video subtitles, or adding transcription to your own software at no cost.
Features
Multilingual recognition: over ninety languages are supported, with accuracy varying by language — the project publishes a word-error-rate breakdown per language for reference.
Speech translation: adding `--task translate` turns non-English speech directly into English text, with the multilingual models (small, medium, large) performing best at it.
Six model sizes: from 39M tiny to 1550M large, needing roughly 1GB to 10GB of VRAM across a tenfold speed range, so you can trade speed against accuracy to fit your hardware; the turbo model is an optimized large-v3 that runs about eight times faster with minimal accuracy loss.
English-only variants: tiny, base, small and medium each ship an `.en` model that outperforms the multilingual model of the same size on purely English audio.
Command line ready: `whisper audio.mp3 --model turbo` transcribes in one command, handling flac, mp3, wav and other common formats, with an option to specify the language and to process several files at once.
Python API: `whisper.load_model()` plus `model.transcribe()` is all it takes to embed transcription into your own program or batch pipeline.
Simple installation: `pip install -U openai-whisper` with ffmpeg available on the system, running on Python 3.8 through 3.11.
Runs locally: inference happens on your machine and audio is never uploaded anywhere, which matters for recorded meetings and interview material.
Open source project: the source is hosted on GitHub under the MIT license, so developers can review the implementation, contribute, or customize it for their own needs.
