1AudioTool runs Whisper on your own machine, biases it with your personal dictionary and your project’s identifiers, then measures whether it actually got better. A 3.1 MB app and a model on your disk — no account, no API key, no upload.
Free dictation handles ordinary sentences well. It falls apart exactly where your day lives: the identifier you type forty times, the surname of the person you are writing to, the second language you switch into mid-sentence. Every correction is retyping — and nothing you correct is ever remembered.
The same failure mode in the audio domain: base.en hears “audio workload… root the PCM”. large-v3-turbo hears “audio worklet… route the PCM”, clean.
You speak English with an accent and your vocabulary is full of filenames, symbols, product names and internal jargon. Whisper is measurably worse on accented speech that people understand perfectly — and worse still on words it has never seen.
Profiles isolate your languages, model, vocabulary and correction history from everyone else's.
You think faster than you type and you want the right result, not a literal transcript: a well-formed prompt, a written message, a note. Rough speech in, finished shape out — reshaped locally before it reaches the clipboard.
Five output modes: Verbatim, Clean, Prompt, Message, Command.
Give a bounded command, hear a concise answer, keep going. Voice actions always show what they parsed; anything consequential needs a keypress. A matching voice is never authorisation.
Four permission levels. Level 3 never runs from speech at all.
You publish in more than one language and you want to keep your own voice and your own source audio. Studio starts from the transcript you already corrected: captions, translation, scripted dub.
Consent-gated voice profiles. No arbitrary third-party cloning, ever.
Audio goes from the system WebView straight to Rust over Tauri IPC, then through local whisper.cpp. There is no Electron, no Node runtime, no localhost server, no API key and no account. Scroll, or drag the stack.
A correction rule firing looks identical to the model genuinely hearing better — unless you separate them. Memory → Recheck re-runs your stored audio against the current dictionary and scores every clip in three columns.
| Column | What it is | What a change means |
|---|---|---|
| R1 raw | What the model heard the first time | The baseline for this clip |
| R2 raw | Same audio, same model, current dictionary, no correction rules | R1 → R2 is real adaptation. The prompt biased the decode and the model heard more. |
| Final | After correction rules run | R2 → Final is text rewritten afterwards. Useful, but not the model. |
Terms are then sorted by what the second round proved about them — because only one of these three verdicts is worth acting on:
The dictionary worked. No rule is carrying this term any more; the decode itself improved.
One rule covers it permanently and your transcript is correct. The model simply never hears the term, and the app says so instead of taking credit.
No deterministic rule can cover an unpredictable mishearing. This is the only category worth your attention. Three occurrences are required before a term is called unstable.
Replay error is edit distance over the longer of reference and hypothesis, not the reference-only denominator. One reference word can expand into five — terminalStore.ts → “terminal storm dot t s” — which scores above 100% under plain WER and lets a single clip dominate the average.
Read a short script covering conversational English, difficult phonemes, names, numbers, URLs, project terms and one code-switched line. You get word error rate, project-term recall, correction burden and latency — then pick a model from what was actually measured, not from your hardware.
⌘1 to ⌘9 follow the rail top to bottom. Capture, knowledge, automation and creation each get their own room, so dictation never turns into a settings hunt.
Correct a transcript to propose a spelling rule for this speaker.
Interface recreated from the app’s own design tokens. Figures shown are example values.
The recorder. Push-to-talk, live transcript, output mode, destination, and the correction you make right after.
What still needs you: clips awaiting review, proposed spelling rules, and imported sources.
Every transcript you have ever dictated — searchable, filterable, replayable, and correctable.
The personal dictionary, and the project packs that make your codebase's vocabulary spellable.
What dictating has actually done for you — counted from what is already on this machine. Free at every tier.
Voice commands with a preview, a permission level, and an audit trail. Voice is never authorisation.
The day, read back to you — written first, so you can check it before you hear it.
Captions, translation and dubbing that start from the transcript you already corrected. The later milestone, shipping in stages.
Models, languages, microphone, storage, shortcuts, appearance, licence and updates.
Items marked PRO need the one-time licence. Everything else is in the free tier, forever.
Categories, not brands — the dictation market is consolidating fast and individual products move every month. Check any specific competitor’s current documentation before you decide.
| Built-in OS dictation | Cloud dictation apps | 1AudioTool | |
|---|---|---|---|
| Where your audio is processed | On-device on supported configurations | Typically uploaded; some offer an offline mode | Always on this machine, at every tier |
| Account required | No | Usually | None. No API key either |
| Knows your project's identifiers | No | A personal dictionary you maintain by hand | A pack built from your repository, gitignore-aware |
| Learns your corrections | No | Often, opaquely | Yes — as rules you can read, scope, merge and forget |
| Proves the learning worked | — | Rarely measurable by the user | Recheck separates model gain from rule gain, per term |
| Two languages in one sentence | No | Varies | Up to three per profile, with a multilingual prompt |
| Voice actions | Fixed navigation commands | Increasingly, via integrations | Four permission levels, preview and audit; Level 3 never voice-only |
| Model choice | Fixed | Vendor's | Three Whisper models on your disk, switched per profile |
| Price shape | Free | Usually a subscription | Free tier forever; Pro is one payment, perpetual |
Local-first is not a setting here; it is the architecture. There is no server to turn off and no key to revoke, because there was never a request to make.
Profiles and transcript history in application-managed SQLite with a write-ahead log. Replayable WAVs as files beside that database. Settings written once per pause, not once per keystroke.
Settings, vocabulary, correction rules, history and transcript/error metrics. It deliberately does not contain microphone audio — and a full data backup never carries your licence key.
Entitlements are verified against a signature, with a 14-day grace window so a transient failure never punishes a customer. Where your audio goes does not depend on what you paid.
Not a subscription, not a seat, not a per-minute credit. A perpetual licence for the machines you name, and a free tier that stays genuinely usable.
Activate in Settings → Licence. Release a device any time. Works offline, with a 14-day grace window.
Download it, dictate for a week, then open Insights and Recheck and see whether it actually learned anything. That is the whole pitch.