Proof
What runs on
your machine.
Kajo Voice runs speech recognition, translation, summarization, and chat on your Mac, Windows, or Linux computer. Imported audio is transcribed by an on-device model (Whisper large-v3 turbo) into a local private library; a single on-device Gemma 4 model then translates, summarizes, and answers questions — all on-device. There is no cloud-processing path: your audio, transcripts, and library never upload.
What never leaves your laptop
Every stage of the loop runs on the machine the audio is already on.
Your machine
No cloud-processing path
-
Drop in
Audio or video, from your Mac, Windows, or Linux machine.
-
Transcribe
Whisper large-v3 turbo writes the transcript and a speaker model labels who spoke — both on that same machine.
-
File
Into kajo.db — your private local library.
-
Ask
Translation, summaries, and cited chat — one on-device Gemma 4 model, over that same library.
The models that run on your machine
Every step of the loop has an on-device model. These are the exact models Kajo runs. None comes in the installer: after you install, the app downloads each one once, or reuses an identical copy already on your computer, and checks it against a pinned SHA-256 checksum before using it.
| Task | Model | License |
|---|---|---|
| Transcription | Whisper large-v3 turbo 99 languages, batch transcription, Mac / Windows / Linux. | MIT |
| Speech detection | Silero VAD v6.2 Skips the silence before transcription, so Whisper only hears speech. | MIT |
| Speaker labels | NVIDIA Nemotron 3 Diarization Finds who spoke when, up to eight speakers per recording, and labels every transcript line. | OpenMDW-1.1 |
| Translation + summaries + chat | Local LLM — Gemma 4 One on-device model translates transcript text, generates summaries, and answers chat questions locally; Gemma 4 is Apache-2.0. It is the largest single piece of Kajo's memory use — about 6 GB loaded, inside a roughly 8 GB total with every model running. There is no cloud option and no separate translation model. | Apache-2.0 |
| Library search (embeddings) | Qwen3 Embedding 0.6B Dense vectors for hybrid search across your archive. | Apache-2.0 |
| Search reranking | BGE Reranker v2 m3 Cross-encoder reranker for cited chat retrieval. | Apache-2.0 |
We track each model's license so you can vet it before deploying on a regulated machine; the third-party notices file installed with the app lists them too, with the license texts. Besides these models, the app ships one small statistical helper: the ELD language detector (Apache-2.0), which helps flag a translated segment for review when it comes out in the wrong language.
How fast it actually is
On the M1 Max these figures come from, a six-hour deposition is about seven minutes of transcription plus under two minutes of speaker labelling. Windows and Linux PCs transcribe and label speakers on the CPU, which takes much longer, and neither has yet been timed on its own hardware. These are our own measurements, from fixtures committed to the repository, on one machine — Apple M1 Max (10-core CPU, 32-core GPU, 32 GB RAM), macOS 26, 2026-09-26. Your times will differ with your hardware, and the honest way to check is the free tier on your own recording.
| Stage | Where | Throughput |
|---|---|---|
| Transcription | Apple Silicon (Metal) Across the committed ASR fixture set — an hour of audio in a little over a minute. | 52.8× realtime |
| Speaker labelling | Apple Silicon (Metal) 624 s of audio in 2.7 s. The 10.5 s sparse-speech fixture sits at the bottom (63 ms) because fixed per-file work weighs more on short files; the 150.6 s deposition runs at 236×. | 167–240× realtime |
| Speaker labelling | Apple Silicon, no GPU The same fixtures with Metal off — the fallback when a Mac cannot use its GPU: 624 s in 7.6 s. The 150.6 s deposition sits at the bottom of the range. | 73–105× realtime |
| Speaker labelling | Windows / Linux (CPU) Two proxies, both run on the benchmark Mac: an Intel macOS build without BLAS — the instruction set and CPU kernels the Windows and Linux x64 builds compile — under Rosetta, 624 s in 83.3 s (7.5×); and the Linux arm64 build itself in a Linux virtual machine, 624 s in 40.4 s (15×). | ~8–15× realtime projected |
On Apple Silicon both stages run on the GPU. Rows marked projected were measured on the same Mac, through Rosetta or in a Linux virtual machine, not on the hardware they describe; we will replace them with direct measurements rather than quiet them.
Frequently asked questions
Does Kajo upload my audio to transcribe it?
No. Transcription runs locally — and so do speaker labels, translation, summaries, search, and chat. There is no cloud-processing path; Kajo never uploads your audio or transcript text.
What does use the network, then?
Never your audio, transcripts, or library. The app uses the network for: the one-time model download (from Kajo's asset CDN, falling back to Hugging Face); licence activation and, once a licence is entered, a check at launch and once a day that it has not been revoked; update checks (on by default — Settings → General); crash reports, only if you agree when the app asks at first run (change it any time with “Send crash reports” in Settings → Privacy), and never with audio or transcript content; usage analytics, only if you tick “Help improve Kajo” (off by default), sent to PostHog in the EU with a random ID created when you opt in; and feedback, only when you send it.
Can I check this myself?
Yes. Once the models are downloaded, turn off your network and use every feature — transcription, speaker labels, translation, search, and chat keep working, because none of them needs the internet. You can also watch the app’s traffic with the network monitor you trust while a recording transcribes.
How is the database protected?
kajo.db is a plaintext SQLite database on your disk — there is no database key, so your OS account and full-disk encryption are what protect it. Turning those on is the single most useful step you can take. Library backups you export are plaintext too, so store them somewhere you trust.
Why does Kajo need 16 GB of RAM?
Because the models run on your machine. Kajo itself uses about 6 GB with the chat model loaded and about 8 GB with everything running at once; the rest of the 16 GB leaves room for your operating system and whatever else you are working in.