Linguist mode turns a podcast, a lecture or a conversation into a live transcript with a translation beside it — transcribed on your own computer, translated by Claude under your own key, saved on your own disk. For interpreters, translators, teachers, and anyone who works or learns between languages.
Status: in KeisenVPA 5.0.0. Linguist is the default mode for new installs; existing clinical installs stay in Healthcare mode. Everything below is built; the hardware report at the bottom is being designed now and is described as such.
Press Start. Speech appears line by line in the language it is spoken, with the translation beneath. Then the transcript becomes a workbench.
OpenAI Whisper runs on your computer. Each pause-cut chunk of speech appears as a line; a grey tail shows what is being said right now and firms up at the next cut. The language is detected automatically.
Claude translates each line as it arrives — text only, never audio, under your own key — into the language you choose. Or translate on your computer into Whisper's rough English, with Whisper's base model (139 MB): if it is not on your computer yet, Start offers to download it and does so once you confirm. Word links and grammar labels still use Claude unless you work offline.
Click a word of the original and the words that translate it light up; click a word of the translation and its source lights up. Separable verbs and multi-word units highlight as one. A word with no counterpart says so.
Add a word and it is labelled: dictionary form, part of speech, grammatical features, gender, and a short learner's note. Export the list as a CSV for Anki.
Option-click (⌥-click) a word of the spoken line, or right-click it and choose Listen, to hear it in context. Listen plays from audio kept in memory, which is gone when the window closes; it never writes audio to disk.
Who said what, computed on your computer. It needs speechbrain and its speaker model, which you install separately: KeisenVPA does not install or download them today, and the window says when they are missing. Nothing about voices leaves your machine.
⌘F finds a word in either language; accents and case are optional.
Show Cyrillic text in a handwriting face, so you learn to read what people actually write, not only what is printed.
Control the podcast app without leaving the transcript (macOS today).
One switch sends nothing to Anthropic from this window. The window lists what works offline (transcription, Whisper's English, speaker labels when installed, the vocabulary list, saving every session) and what needs a connection (Claude translation, word links). Grammar labels for words you add offline are filled in, and their sentences sent, once you switch it off or reconnect. Working offline translates with Whisper's English, which uses Whisper's base model; if it is missing, Start offers to download it, from OpenAI's public model server rather than Anthropic, once you confirm. A dropped connection is noticed and the window says so; with the base model present it switches to Whisper's English by itself.
Every line and every translation is written to disk the moment it arrives, so a crash never loses translations you paid for. Save… writes a named copy wherever you like. The audio is saved too only if you turn on Save audio (off by default): then each session's audio is kept beside its transcript as a FLAC file, and the window says so while it records.
The footer estimates, on your computer, what this session's Claude usage costs on your key. Keisen does not meter it.
The microphone, or a virtual audio device that carries your podcast app's sound. Every input is listed, virtual ones included.
Whisper runs on your computer. No audio is sent anywhere, ever. Its transcription model (turbo) is downloaded after you confirm.
Claude translates the text under your own Anthropic key, a few sentences at a time with the preceding lines for context. Offline, Whisper's English stands in when its base model is on your computer; that model, too, is downloaded after you confirm.
Word links, vocabulary, listen-again, search. The session is already on your disk; export the vocabulary to Anki when you are done.
The full statement is in the privacy policy, Section 3.3. In short:
.pcm.partial file and converted when you stop, or the next time the window opens if the app quit first. Moments muted while a Listen clip plays are saved as silence.~/Library/Application Support/KeisenVPA/Live Translation/ (not synced to iCloud); on Linux: ~/.local/share/KeisenVPA/Live Translation/. Readable only by your user account. The window links to the folder. Delete the files whenever you like.Linguist mode has its own bundled package, KeisenVPA Linguist Templates 1.0.0; the clinical templates do not appear in this mode. The templates are buttons in the main window's chat, behind the Live Translation window. They work on whatever is in the conversation: pasted or typed text, a Live Translation session file you attach, dictation, or a screenshot. Any language, explained in English unless you ask otherwise, with stress marks for Russian. A point the model is unsure of is marked "(unsure)" and listed at the end instead of being invented. Each one sends the material to Claude under your key, like any other message.
Sentence by sentence: a gloss, a natural translation and the grammar you need to read it, then the vocabulary worth learning.
LinguistA literal and a natural translation, with notes on idiom, register and false friends.
LinguistFrom a Live Translation session file: a summary, a check of the transcription, key vocabulary, grammar points and five comprehension questions.
LinguistYour text corrected, every correction explained at your level, and what to practise next.
LinguistAn Anki import file (tab-separated: front, back, notes) from the words in the material.
LinguistTen questions from the material (vocabulary, grammar, comprehension), with the answers hidden at the end.
LinguistOne word in depth: every form, the aspect pair, related words, collocations and examples. Type the word first.
LinguistReal-time transcription and translation need more machine than most laptops have. This page will tell you honestly which tier yours is in, from measurements — not from assumptions.
So far we have measured two computers, both the developer's: an Apple M4 Max, which keeps up with speech with every feature on, and a Framework Laptop 13 (Intel i7-1370P, no GPU, Linux), where transcription runs about 30 seconds behind live once the spoken language is set in the From: box (with automatic detection it fell as far as about 90 seconds behind before it began to catch up). Claude translation was not measured on that laptop; it depends on the network rather than the processor, and took about 1.5 seconds a line on the Mac. We will not extrapolate from either to yours.
So the app should not assume either. We are designing a capability report that measures your computer, shows the lag, and turns down extras before audio is ever dropped. It is not in the app yet. Today a slow computer shows “Whisper is behind” or “Whisper is slow — live tail paused”, and audio can be lost when it falls far behind.
CPU-only machines, including Linux laptops without a GPU, are a first-class target: they transcribe with a lag, and the design is meant to measure them, not to turn them away. Linguist mode runs on Apple Silicon Macs. On Linux it runs from source today; the packaged Linux build does not include it yet.
| Tier | What it means for you | Measured on |
|---|---|---|
| Live | Transcript and translation keep up with speech. Word links, speaker labels and vocabulary work as you listen. | Apple M4 Max (developer's): keeps up with every feature on |
| With a lag | Lines arrive behind the speaker. On a CPU the app turns off word timing, keeps the grey live line off until the language is set, and pauses it whenever Whisper runs slow — on the laptop measured, most of the time even with the language set. It says “Whisper is behind” or “Whisper is slow — live tail paused” rather than showing the lag in seconds. | Framework Laptop 13, Intel i7-1370P, no GPU, Linux (developer's): transcription about 30 seconds behind live with the language set in the From: box, up to about 90 seconds behind with automatic detection. Claude translation not measured there (network-bound; about 1.5 s a line, measured on the Mac) |
| Afterwards | Planned, not in the app yet: the session would be transcribed and translated when you stop, not while you listen, with word links, vocabulary and search working the same once it is done. | To be filled from measurements |
Placeholder, on purpose. This table is filled only from measurements, with the machine named, and says so where something was not measured; the "Afterwards" row waits for one. The in-app capability report that will tell you your own machine's tier on first start is still being designed.
KeisenVPA is delivered directly today, on your own Anthropic API key. Linguist mode runs on Apple Silicon Macs; on Linux it runs from source today, and the packaged Linux build does not include it yet. Tell us which platform you have.