Speech to Text

Dictation as a writing method

Speaking is roughly three times faster than typing — 130–150 words per minute against 40–50 for an average typist — which makes dictation a serious drafting tool, not just an accessibility feature. The technique that works: dictate a messy first draft without stopping to fix errors, then do a typed editing pass. Trying to dictate perfect sentences defeats the speed advantage. Writers use this for first drafts, professionals for meeting follow-up emails dictated immediately while context is fresh, and students for converting spoken study notes into written ones.

The local-AI trade-off, honestly

This tool runs a neural speech model entirely in your browser via WebAssembly. The privacy benefit is absolute — audio of your voice never leaves the device, which matters for legal dictation, medical notes, and anything covered by confidentiality. The trade-offs are equally real: the model downloads once (tens of megabytes, then cached), transcription speed depends on your hardware, and the largest cloud models still lead on heavily accented or noisy audio. For everyday dictation on a modern laptop, the gap is small; for archival transcription of difficult recordings, plan a manual review pass either way.

A realistic post-transcription workflow

Raw transcription output needs the same treatment regardless of which engine produced it:

1. Punctuation and paragraph pass — spoken language runs on; break it into written-language sentences.
2. Names and terms pass — search for the proper nouns you know appeared and fix their spellings consistently.
3. Filler removal — delete the “um”, “you know”, and false starts (or keep them, if you’re producing a verbatim record).
4. Structure pass — for meeting notes, reorganize chronological talk into decisions, actions, and open questions.

Budget roughly 2–3 minutes of editing per minute of audio for publishable text; transcription gets you 90% of the way, editing is the irreducible 10%.