Short answer: Whisper is OpenAI's open-source speech recognition model. Its code and model weights are released under the MIT License, so the model itself is free; what costs money is the computer that runs it. You can use it free online (SpeakScribe runs Whisper large-v3-turbo; a free account transcribes 120 minutes a month), or install it yourself if you have a recent graphics card. How accurate it is depends almost entirely on the language: in our test of 40 languages, 20 came out nearly word-perfect and 8 were not usable.
What Whisper is
Whisper turns speech into text. It was trained on a large amount of multilingual audio, recognises about 99 languages and does not need to be trained on your voice. OpenAI publishes several sizes of the model; bigger ones are more accurate and slower. These are the figures from the official Whisper repository:
| Model | Parameters | Graphics memory needed | Relative speed |
|---|---|---|---|
| tiny | 39 M | ~1 GB | ~10× |
| base | 74 M | ~1 GB | ~7× |
| small | 244 M | ~2 GB | ~4× |
| medium | 769 M | ~5 GB | ~2× |
| large | 1550 M | ~10 GB | 1× |
| turbo | 809 M | ~6 GB | ~8× |
Relative speed as published by OpenAI (English speech on an A100 GPU); real-world speed varies with hardware and language.
The turbo model is an optimised version of large-v3: much faster, with a small loss of accuracy. One catch: OpenAI notes that turbo is not trained for translation and returns the original language even if you ask it to translate. For translation into English, the repository recommends the medium or large model.
How accurate Whisper is: 40 languages measured
In October 2026 we ran Whisper large-v3-turbo (the same call SpeakScribe uses) on 800 sentences from Google's FLEURS test set: 20 recorded sentences in each of 40 languages, read by native speakers, about 165 minutes of audio. We set the language explicitly and compared the output with the correct text. The groups below use the median character error rate per language:
| Result | Languages | What it means |
|---|---|---|
| Quick read-through (≤ 4% of characters wrong) | 20, including English (2.6%), Spanish and Italian (0.0%), German, French, Portuguese, Russian, Arabic (2.8%), Japanese (2.4%), Chinese, Korean, Turkish, Vietnamese, Indonesian | Most sentences come out right; check names and numbers. |
| Light editing (≤ 8%) | Persian (7.9%), Swahili | A few fixes per paragraph. |
| Moderate editing (≤ 16%) | Urdu (8.1%), Hindi (9.7%), Hebrew, Thai, Kannada | A solid draft; something to fix in most sentences. |
| Heavy editing (≤ 40%) | Tamil (22.2%), Telugu (29.5%), Bengali (38.6%), Marathi, Nepali | Saves typing, but check every sentence. |
| Not usable | Amharic, Burmese, Gujarati, Malayalam, Pashto, Punjabi, Somali, Uzbek | Most of the text comes out wrong. |
The full table, with word error rates and an example sentence for every language, is on audio to text; the method is on how we test.
The pattern is clear: Whisper is excellent in widely written languages with plenty of training audio and weak in many South Asian and African languages. In the "not usable" group all 20 Amharic sentences and all 20 Burmese sentences came out more than half wrong: Amharic as real-looking Amharic words with the wrong meaning, Burmese as a jumble of characters from other scripts. If your language is in the bottom two groups, read every line of the transcript before you rely on it.
Using Whisper free online
- No install, no graphics card. Upload a file in the browser and get text, a timestamped version and SRT/VTT subtitles.
- Limits apply. On SpeakScribe a free account transcribes 120 minutes a month, up to 30 minutes per file; you can also try 20 minutes without an account.
- It is fast. Across 2,140 real transcriptions in October 2026, the median file under 5 minutes was done in about 1 second of processing and files of 5–15 minutes in about 2 seconds, because the model runs on dedicated hardware.
- Your file leaves your computer. Check how long the service keeps it. On SpeakScribe, files are deleted automatically when they expire (after 7 days on the free plan).
Running Whisper yourself
- What you need: Python, the command-line tool ffmpeg, and
pip install -U openai-whisper. Thenwhisper audio.mp3 --model turbo --language Englishwrites the transcript next to the file. - Hardware decides the experience. The turbo model needs about 6 GB of graphics memory and large about 10 GB. Without a suitable graphics card it runs on the processor, which works but can take longer than the recording itself.
- Privacy and cost. Nothing leaves your computer and there are no monthly limits, which suits confidential recordings or very large archives.
- What you get: the command writes plain text and subtitle files (SRT, VTT) next to the recording. Speaker names, spelling fixes and translation are up to you.
| Free online | On your own computer | |
|---|---|---|
| Setup | None | Python, ffmpeg, a model download |
| Hardware | Any browser | A graphics card with 6 GB+ for turbo |
| Limits | Monthly minutes and file length | None |
| Privacy | File uploaded to the service | Stays on your machine |
| Best for | Occasional files, phones, quick subtitles | Large archives, confidential audio |
Getting a better Whisper transcript
- Choose the language instead of relying on auto-detect, especially for short clips or accents. All our measurements set it explicitly.
- Record close to the microphone. Echo and background noise cost more accuracy than any model setting (see how to record clear audio on your phone).
- Trim long silences. Like other speech models, Whisper can write text where nobody speaks or repeat a line during long quiet stretches.
- Read names and numbers. These are the most common errors even in the best-scoring languages.
FAQ
Is Whisper AI free?
The model is free and open source under the MIT License. Running it costs hardware or a service: free online tools give you a monthly allowance, and on your own computer you need a capable graphics card for good speed.
How accurate is Whisper?
It depends on the language. In our test of 800 sentences in 40 languages, 20 languages had 4% or fewer characters wrong (English 2.6%), while 8, including Amharic, Burmese and Punjabi, were not usable.
Is Whisper accurate for Hindi, Urdu and Bengali?
Partly. Urdu (8.1% of characters wrong) and Hindi (9.7%) give a solid draft that needs editing; Bengali (38.6%) needs heavy correction.
Can Whisper translate audio into English?
The multilingual models can, but OpenAI says the fast turbo model is not trained for translation. SpeakScribe transcribes first and then translates the text, see translate audio.
Does Whisper label different speakers?
No. Whisper writes what was said, not who said it. Add speaker names while you review a timestamped transcript.