Home → Blog → Which Whisper model to choose for speech recognition
Which Whisper model to choose for speech recognition
Whisper ships in several sizes, from tiny at 39 million parameters to large at 1.5 billion. The gap is not only accuracy: it decides speed, VRAM and ultimately the cost of a minute of audio. Here is how to pick one for the job.
The line-up
| Model | Parameters | VRAM | Speed | Good for |
|---|---|---|---|---|
| tiny | 39M | ~1 GB | very fast | rough drafts, voice commands |
| base | 74M | ~1 GB | fast | short clips with clean audio |
| small | 244M | ~2 GB | moderate | clear speech, single language |
| medium | 769M | ~5 GB | slower | noise, accents, mixed languages |
| large-v3 | 1550M | ~10 GB | slow | maximum accuracy, hard audio |
There is also large-v3-turbo: the same large model with a trimmed decoder. It runs several times faster and gives up almost nothing on ordinary speech.
Where small models break
On a studio recording of a single speaker the sizes are nearly indistinguishable. The gap opens up exactly where audio gets hard:
- Proper nouns and jargon. A small model confidently turns a surname into a similar everyday word.
- Accents and fast speech. tiny and base drop chunks of sentences when speakers overlap or swallow endings.
- Language switching mid-recording. Small models handle it poorly and sometimes translate instead of transcribing.
- Noise. Street, car, open office — that is where a small model starts inventing text.
What the choice costs
Your own GPU for large-v3 means 10 GB of VRAM and a machine permanently tied up. Rented cloud GPUs bill by the hour, not by the minute of audio, so idle time costs exactly as much as work. A per-minute service removes the question: the big model always runs and you pay for actual volume.
Rule of thumb: below roughly ten hours of audio per day, owning a GPU costs more than paying per minute.
What to pick in practice
For voice messages, calls and meetings take the largest model available: the accuracy gain outweighs the speed loss, and fixing errors by hand is more expensive than either. Small models earn their place in two cases — on-device recognition without a network, and rough passes over huge archives where volume matters more than every word.
Our API runs large-v3-turbo, so there is nothing to choose: requests pass whisper-1 as the model name purely for compatibility with existing code.
Try it on your own recordings. Sign-up takes a minute, and the free minutes are enough to judge the quality.
Get a free API keyFrequently asked questions
How does large-v3-turbo differ from large-v3?
Turbo has a shortened decoder: several times faster with nearly identical accuracy on ordinary speech. On difficult audio the full large model still wins by a fraction of a percent.
Do I have to specify a model in the request?
Pass whisper-1 — the name exists for compatibility with existing code. Which model runs on our side is not something you need to configure.
Is the biggest model always more accurate?
On hard audio, yes. On a clean single-speaker recording the difference between small and large is a fraction of a percent and invisible in practice.
How much VRAM does a self-hosted setup need?
Around 10 GB in fp16 for large-v3. small and base fit into 2 GB, with a noticeable accuracy penalty.
Related reading
- Self-hosted Whisper or a transcription API: what costs less — An honest comparison of running Whisper yourself versus using an API: GPU cost, maintenance, time to first result and the volume where self-hosting pays off.
- How to improve speech-to-text accuracy: eight practical fixes — Eight changes that actually move transcription quality: audio preparation, language hints, the prompt parameter, chunking, formats, and how to measure word error rate.
- Which languages speech recognition actually handles well — The 99 supported languages are not equal: which are reliable, where accuracy drops, how detection works, and what to do with recordings that switch languages.