Speech to Text

Why local transcription matters here

Consider what actually gets transcribed: interviews with sources, medical consultations, legal depositions, therapy sessions, internal meetings, research participants, family recordings. It is one of the most consistently sensitive categories of file anyone handles.

Every mainstream transcription service works by uploading that recording to their servers. Most are reputable, but the recording has still left your control, is subject to a retention policy you did not negotiate, and in many professional contexts — journalism protecting a source, a clinician handling patient data, a lawyer with privileged material — that transfer is itself the problem, regardless of what happens next.

Whisper is an open speech recognition model, and transformers.js can run it inside a browser tab. That means the audio is decoded, processed and transcribed on your own processor. There is no upload, no account, no per-minute charge, and no copy anywhere else.

Choosing a model size

Tiny is about 40 MB and transcribes several times faster than real time on a modern laptop. For clear speech — a single speaker, decent microphone, no background noise — it is often good enough, and it is the right place to start.

Base roughly doubles the size and noticeably improves accuracy on accented speech and slightly noisy recordings. Small is around 250 MB and is meaningfully better again, particularly on technical vocabulary and proper nouns, but is slow enough that a long recording becomes a genuine wait.

The English-only variants are more accurate than the multilingual ones at the same size, because the whole capacity is spent on one language. Use them unless you actually need another language.

Each model downloads once and is then cached by your browser, so the cost is paid on the first use rather than every time.

What affects accuracy

Audio quality dominates, in the same way scan quality dominates OCR. A close microphone in a quiet room produces dramatically better results than a phone on a table across the room.

Overlapping speech is the hardest case. Whisper transcribes a single stream and does not separate speakers, so crosstalk in a meeting produces a jumbled passage. It also does not label who is speaking — speaker diarisation is a separate technology this does not include.

Proper nouns, product names, acronyms and technical jargon are the most common errors, because the model resolves ambiguous audio toward common words. A name it has never encountered will usually be rendered as the nearest ordinary word.

Long silences and music can produce repeated or hallucinated phrases — a known Whisper behaviour where the model fills a gap with plausible text. If you see a sentence repeating, that is what happened.

Realistic expectations on speed

Running a neural network in a browser tab uses the CPU through WebAssembly rather than a dedicated accelerator, so it is far slower than a server with a GPU. Budget roughly real-time or a little better with the Tiny model, and several times slower than real time with Small.

A one-hour recording on the Small model is not a practical browser task. For anything long, use Tiny or Base, or split the recording and transcribe it in sections.

Keep the tab in the foreground — browsers throttle background tabs aggressively, and a backgrounded transcription can slow to a crawl.

The SRT export makes this useful beyond transcription: run it against a video and you have a starting set of subtitles, which the subtitle converter on this site can then retime or reformat.