Whisper: transcription that handles accents
Open-source speech recognition you can run locally or through an API, and the accuracy limits worth knowing.
Pricing
Free to run locally; hosted APIs billed per minute (check the official page)
Official site
Open the official website (opens in a new tab)
Prices and plans change. Check the official page before you pay.
Whisper is an open-source speech recognition model from OpenAI. It transcribes and translates audio in dozens of languages, and it handles accents, background noise and imperfect recordings better than the tools that came before it.
What it is good at
- Accented and noisy speech. It was trained on a very large, varied set of audio, which shows in real-world recordings.
- Translation. It can transcribe non-English audio directly into English, which is useful for research interviews.
- Timestamps. Word and segment timings make subtitles practical.
- Running locally. The weights are public, so a confidential recording never leaves your machine.
Where it fails
- Names and jargon. Proper nouns, product names and technical terms are the usual errors; a glossary or a manual pass fixes most of them.
- Hallucination in silence. On long silent passages it can invent text, a known failure that matters for meeting recordings.
- Speaker labels. It does not identify who spoke; diarisation is a separate step.
- Very long files need chunking, and the seams can lose a word.
Running it locally
The reference implementation needs Python and ffmpeg. faster-whisper is a reimplementation that runs several times quicker on the same hardware, and desktop apps wrap it for people who do not want a terminal.
Model size is the main trade: small is fast and adequate for clear speech, large is slower and noticeably better on accents and noise. A modern laptop handles small comfortably; large wants a GPU.
Accuracy in practice
Expect a few errors per minute on clean speech and more on a phone call. The realistic workflow is machine transcription plus a human pass, which is still an order of magnitude faster than typing from scratch.
Always spot-check names, numbers and anything that will be quoted.
Pricing
Running the model locally is free. Hosted transcription APIs charge by the minute, typically a fraction of a cent, and bundled cloud services add storage and processing. Compare before transcribing hundreds of hours.
Who it is for
Journalists, researchers, podcasters and anyone who needs searchable text from audio. It is also the standard first step for building subtitles.
Alternatives
AssemblyAI and Deepgram offer hosted transcription with speaker labels and better real-time support. YouTube and Zoom generate transcripts automatically and less accurately. For a confidential recording, local Whisper is the safest choice.