Skip to main content

Architecture honesty

Why on-device transcription beats cloud “cleanup”

Cloud vendors sell upload-then-clean pipelines. Kajo transcribes on-device with two decode profiles — speech, and one for sung vocals — chosen automatically. No audio leaves the device.

Kajo Voice on-device transcript detail without cloud cleanup
Transcription finishes on your hardware — no cloud cleanup step.

The cloud cleanup pitch

Once a vendor has your file, they can run proprietary filters, charge for “enhanced” accuracy, and keep a copy under their retention policy. The marketing language is cleanup; the architecture is custody. For interviews, depositions, session recordings, and source tapes, that trade is often unacceptable.

Browser-based tools compound the issue: your recording traverses their network before you see a transcript. Even if the vendor is well-intentioned, you have expanded the trust boundary to include their staff, subprocessors, and legal process.

What Kajo actually runs

Kajo Voice transcribes on-device on your Mac, Windows, or Linux machine: one engine with two decode profiles — speech, and one for sung vocals — chosen automatically before transcription, plus a small duplicate-line cleanup after decoding. No open-ended re-transcribe, no cloud “cleanup” stage. Translation, summaries, and cited chat run on the same on-device Gemma 4 model.

That is intentional simplicity. Upload-then-clean pipelines sound impressive in a blog post; they also hide failure modes and invite paid cloud fallbacks. Kajo has no cloud fallback of any kind. If the model misses a word, you hear the same audio the model heard — on your disk.

  • Audio never uploads — there is no off-device cleanup stage to sell.
  • Two decode profiles for conflicting audio (speech versus singing) — still fully local.
  • Free tier lets you validate accuracy on your accents and mics before you pay once.

How to get better transcripts without a cloud pipeline

Capture quality still matters. Prefer a close mic over a room mic. Reduce HVAC and keyboard noise at the source. For music and songs, leave Auto on or pick Lyrics so silence detection does not strip sung vocals. Import existing recordings into the library and ask them with citations — the archive compounds even when a single file is imperfect.

If a cloud vendor’s cleanup demo looks magical, ask where the audio went, how long they keep it, and what happens under a subpoena. Kajo’s answer is short: it never left your machine.

Try the on-device path

Download free, transcribe a representative clip, then unlock Lifetime when the allowance is not enough — one payment, local-only.

Frequently asked questions

Why do cloud tools advertise audio cleanup?

Once your audio is on their servers, vendors can run proprietary post-processing and bill for it. That convenience is also custody: the recording left your machine. Kajo never uploads audio, so there is no cloud cleanup stage to sell — and no third-party custodian.

How should I improve accuracy with Kajo?

Use a decent microphone, reduce room noise at the source, and test the free tier on a representative clip before unlocking Lifetime. For songs, leave Auto on or pick the Lyrics profile. After decode, fix stubborn names once with inline segment edit — that does not replace a clean capture.