The Audio to Text tool and the captions in the video editor run speech recognition inside your browser. Nothing is uploaded: the model comes to your device instead. That makes the choice of model a practical one. A bigger model gets more words right, but it is a bigger download and does more work for every second of sound.
The model is Whisper, released by OpenAI as open weights, in versions converted to run in a browser. The site offers four sizes. Here is what each costs and when each is worth it.
The four sizes and what they download
These are the download sizes the tools show before anything is fetched, in decimal megabytes as your browser and phone report them:
| Model | Audio to Text, on the processor | Audio to Text, on the graphics chip | Video captions (word timing) |
|---|---|---|---|
| Tiny | 44 MB | 122 MB | 44 MB |
| Base | 80 MB | 209 MB | 80 MB |
| Small | 252 MB | 589 MB | 252 MB |
| Large (large-v3-turbo) | not offered | not offered | 566 MB, graphics chip only |
Why the same model is bigger on the graphics chip: on the processor the tools use 8-bit versions of the weights, which are small and fast on a CPU. On the graphics chip, Audio to Text uses a full-precision encoder, which graphics chips run well, and so the file is larger. Large uses 4-bit weights with half-precision maths, which is why it fits in 566 MB at all, and why it needs a graphics chip that supports half-precision.
Speed against accuracy
Each step up is slower. The tools describe Base as noticeably more accurate than Tiny and about twice as slow, and Small as the most accurate of the three and the best for Hindi and other Indian languages. As a rough sense of speed, the site measured Tiny timing every word of a 14-second recording in about 7 seconds on a laptop processor.
On clear speech in English, the gap between sizes is small enough that Tiny or Base is often fine and saves a long download. The gap opens up in three situations:
- Indian languages. Tiny and Base are weak on Hindi and the other Indian languages. When you choose one of them, or the tool detects one, it recommends Small.
- Music. The smaller models make words up over music. In the site's own test on a Hindi song with loud music, Base produced almost nothing usable, Small got about half the words recognisable, and Large got most lines right. Songs are covered in detail in auto captions for songs.
- Poor recordings. Background noise, distance from the microphone and several people talking at once all favour a larger model.
A model that gets more words right also saves time after the run. Correcting a wrong word by hand takes far longer than the extra seconds the larger model needed to get it right, so for anything you will publish, spend the time on the model rather than on the edits.
Graphics chip or processor
When you open the tool, it asks your browser whether WebGPU is available, which is the way a web page can use your graphics chip. If it is, Audio to Text runs the model there, and the tool tells you so: on a graphics chip it is usually faster than the recording plays. If the model cannot start on the graphics chip, the tool falls back to the processor by itself.
On the processor, the tool's own guidance is that on a computer Tiny and Base usually keep up with the recording, and that on a phone you should allow longer and use Tiny.
The video editor's captions work differently. They need a time for every single word, so that the word being sung or spoken can light up, and that timing step runs on the processor anyway. Tiny, Base and Small for captions therefore always run on the processor, and the same file works on every computer and phone. Large is the exception: it runs only on a graphics chip, and it is offered only when yours can run it. If it fails to start, the editor says so and picks Small instead.
WebGPU support depends on your browser, operating system and graphics driver, so the same laptop can offer Large in one browser and not in another.
Phones and laptops
On a phone the tools start on Tiny, and the video editor does not offer Large. There are three reasons. Small is a 252 MB download, which is a lot on mobile data. Phones have less memory to hold a model and a long recording at once. And a phone processor is slower, so a long recording on Small can take a very long time.
You can still choose Base or Small on a phone if you need the accuracy, for a short recording on Wi-Fi. For an hour-long interview, or a song, a computer is the better tool. Audio to Text accepts recordings up to 3 hours, but how long that takes depends entirely on the model and the device.
Why the first run downloads, and the second does not
The first time you use a model, the button says so: "Download model (252 MB) and transcribe", for example. The files come from Hugging Face, where the converted models are published, straight to your browser. Your browser then keeps them in its own storage for this site. The next time, the tool checks that every file is already there and starts without downloading anything.
A few things to know about that stored copy:
- Each model, and each version of it (processor or graphics chip, plain or word-timed), is its own set of files. Trying Small after Base means a second download.
- Clearing this site's data in your browser settings removes the models. So does a private window, which forgets everything when it closes.
- Cancelling stops a download or a run in progress. A model that had finished downloading stays stored.
- The download is the model, never your recording. Your audio stays on your device the whole time.
Two settings that matter as much as the model
Choose the language. Detection listens to up to three 30-second stretches of the recording (where the sound starts, the middle and three quarters in) and guesses. Choosing it is faster and more accurate, and on a mixed or noisy recording it avoids a wrong guess that colours the whole transcript.
Paste the words if you have them. If you already have a script or lyrics, both tools can time your text to the recording instead of transcribing it. Your words are kept exactly; the model only finds when each line is said. That uses the word-timing version of the model, on the processor. It is the most reliable route to correct captions whatever size you choose.
Which one to pick
| Situation | Model |
|---|---|
| Clear English speech, phone, or a quick check | Tiny |
| Clear speech on a computer, meeting notes, a lecture | Base |
| Hindi or another Indian language, noisy audio, anything you will publish | Small |
| Songs and Hindi captions, on a computer whose graphics chip can run it | Large |
| You already have the words | Paste them; any size |
Whatever you choose, read the result before you use it. Names, numbers and technical terms are where every size makes mistakes, and the tools mark the parts they could not write rather than hiding them. Once the text is right, SRT, VTT or burned-in captions covers how to deliver it.