K—Speech · speech program

Nordic speech models, built from the ground up.

Nordic speechmodels, built fromthe ground up.

K—Speech is our proprietary speech program. We build the models, training data and evaluation loop ourselves, keeping the capability in-house.

K—Speech · evaluation workspace

Nordic speech, reviewed by people.

The workspace below is part of K—Speech. Every take passes a human before it counts: play it, mark the exact second, leave a note, approve it or send it back.

Become a reviewer Open the workspace
Inside the K—Speech evaluation workspace: press play, hit Edit, drag across the wave to slice it, tag it and leave the note.
console.kapllanlabs.io / human-review / blind-en
How review works

Four passes before a take counts. Nothing ships on a model's word alone.

Blind listen
The reviewer hears the take without knowing which system produced it. Judgement first, provenance after.
Removes the pull toward a favourite model.
Mark the second
Problems are pinned to a region on the wave, not to a star rating. A slice, a timestamp, a sentence about what is wrong.
Every note is addressable in training.
Second opinion
Anything marked goes to a second reviewer before it is accepted or sent back for a re-record.
Disagreement is data, not noise.
Back into training
Approved takes and their notes return to the corpus with the reason attached, so the next run knows what failed.
The loop is the point.
One model family

One model family. Every function of speech.

Delayed Streams Modeling is symmetric: the same architecture, codec and aligner run both directions. Swap the delay and you swap the task. Everything below shares one Norwegian/Swedish adaptation recipe.

Speech to text ASR

audio in
text out
·TogettilOslogårfraspor
text stream delayed · 0.5s
⇄

Text to speech TTS

text in
Heipådegdetgårbra··
audio out
··
audio stream delayed · 1.28s

Same architecture, both directions. Given text and audio streams, ASR corresponds to the text stream being delayed, while the opposite gives a TTS model.Zeghidour et al. · arXiv:2509.08753

In the console

Everything a reviewer needs, and nothing else.

Mark the second that is wrong
Drag across the wave, tag the fault, and the note lands on the exact region of the take.
Region tagged
00:14–00:17 Sibilance00:31 Breath
astrid_prompt_01
Nordic voice · ckpt_42000
3 notes
123
00:06 Clip00:14–00:17 Sibilance00:31 Breath
Queue management
Every take, its state and its notes in one queue: unreviewed, approved, rejected, on hold.
sigrid_prompt_14
Nordic voice · ckpt_42000
Loop
00:14–00:17 Sibilance00:31 BreathApproved
Share this review
×
https://speech.kapllanlabs.io/r/astrid-01
Copy
Send it to a linguist
A read only link to the take and the notes. No account, no export, nothing to install.

Speech is judged, not scored.

We open the review console to partners who record with us. If your organisation has Nordic audio and a standard to hold, we would like to hear from you.

Become a reviewer → hello@kapllan.ai →