TranscribeLocally

How accurate is browser transcription, really?

Most transcription tools answer this question with a percentage and no context — “99% accurate!” — which is meaningless without knowing on what audio, measured how. Here is the useful version.

The short answer

On clear, close-microphone English speech, the transcript you get here is good enough that you will spend your editing time on proper nouns and specialist vocabulary rather than on ordinary words. Dictation, a podcast recorded on decent microphones, a one-to-one interview in a quiet room: all fine.

On hard audio, it is meaningfully worse than the paid cloud services. Several people talking over each other, a phone in the middle of a conference table, a strong accent the model saw little of, background music, a windy street: expect real errors, and expect them clustered exactly where the content is most interesting.

That gap is not a bug in this implementation. It is the direct consequence of two deliberate choices, both of which you can see and neither of which is hidden.

Why the cloud services are better on difficult recordings

They run much larger models. Whisper comes in sizes. This site ships base (74 million parameters) and tiny (39 million). The full-size large checkpoint is around 1.5 billion — twenty times bigger — and it is dramatically better at exactly the hard cases: accents, noise, overlapping speech. Cloud services run that one, or something comparable, on a GPU in a datacentre.

Nobody is going to download several gigabytes of weights to transcribe a voice note in a browser tab. So the honest position is that we ship what is reasonable to download, and we tell you what that costs.

The models here are quantized. The weights have been compressed from 32-bit floating point to 8-bit integers, which makes the download about four times smaller and the arithmetic faster. It also loses a little precision. The effect is small — usually invisible on easy audio — but it is not zero, and on marginal audio it can be the difference between a word being right and being plausible.

What “accuracy” even means

The standard measure is word error rate: the proportion of words inserted, deleted or substituted compared with a correct human transcript. A 10% WER means one word in ten is wrong, which sounds tolerable and reads much worse than it sounds — errors are not evenly distributed.

Published WER figures are nearly always measured on benchmark datasets of read speech, and they are usually quoted for the largest models. They are close to useless for predicting what your recording will do, because your recording is not clean read speech, and this is not the large model. That is why you will not find a WER number on this site: quoting someone else’s benchmark for a different model as if it described this one would be dishonest, and measuring our own properly is a research project we have not done.

What matters far more than any number: the same model on the same content will vary enormously with the recording quality. Microphone distance, room reverberation and crosstalk swamp every other variable, including model size. A tiny model on a headset recording beats a large model on a phone across the room.

The specific ways it goes wrong

Knowing the failure modes makes proofreading much faster, because you know where to look.

Proper nouns. Names of people, companies, places and products are the single largest error category. The model substitutes a plausible common word — and it does so fluently, so nothing looks wrong. These repeat, which is the saving grace: fix the first occurrence, then use the search box to find and correct the rest in a minute.

Numbers and units. Doses, dates, prices, case references, phone numbers. Always check these against the audio; never assume.

Crosstalk. Where two people speak at once, the model outputs one voice, or a blend, with no indication that anything was lost. This is the failure mode people most often miss, because the result reads perfectly well.

Hallucination on silence and music. Whisper is trained to produce speech, so given a long silence or a musical passage it sometimes invents a plausible sentence — classically a repeated phrase, or something that sounds like a YouTube sign-off. This tool automatically strips runs of identical repeated segments, which removes the most common form. Anything left tends to sit at obvious boundaries and takes seconds to delete.

Punctuation and paragraph breaks are predictions, not transcriptions. A speaker who runs sentences together will get long sentences. Nothing is wrong; that is what the speech sounded like.

What it does not do at all

It does not identify speakers. That requires a separate model — diarization — which Whisper does not include. Tools that show “Speaker 1 / Speaker 2” are running one, and on similar voices or any overlap it produces confident, wrong attributions. In an interview or investigation transcript, a misattributed quote is worse than no attribution at all, so this site asks you to label speakers manually instead. It takes a handful of clicks and it is right.

It does not know your vocabulary. Cloud services often let you supply a custom word list — participant names, drug names, product names — which measurably helps. There is no equivalent here. Search-and-replace after the fact gets you most of the way.

It is not a certified transcript. For anything that goes on a court record or into a clinical note, a machine draft — from any tool — is a starting point that a human must verify.

Choosing between the tiers

Switching tiers costs one click and a download you only do once.

When you should use something else

Be clear-eyed about this:

The case for doing it here is when the recording is one you should not hand to anybody else. Then the comparison is not against a cloud service; it is against typing it yourself.