Speech recognition for Amedia

Nynorsk · Northern Sámi · Kven

Evidence snapshot ·

Nynorsk has a measured baseline. Northern Sámi has specialist models worth testing. Kven needs authorised evaluation data before we can make a defensible performance claim.

These results describe separate tests, not the current transcription demo. They are not a cross-language ranking.

Nynorsk

Measured locally

Norwegian speech
to Nynorsk text

15.64% WER1

NB Samtale · conversation

10.89% WER · Common Voice read speech

Scored: 246 conversation clips;
322 read-speech clips.

Limit: Short-clip development tests, not newsroom validation or proof of best overall performance. Speaker coverage is limited. We do not know whether training included this material.

Next: Test the same setup on representative newsroom audio with permission and editor review.

Northern
Sámi

Published evidence

Lule and Southern Sámi
not assessed

32.3% WER2

2025 study · 1-hour test

Read and spontaneous speech.
A published result, not our own test.

Limit: Newer specialist models need testing.3 We have no local results. Access to evaluation audio is limited.

Next: Obtain audio with permission and fluent reviewers. Compare models on the same test.

Kven

Not established

Kven-specific evidence
required

No verified score to quote

In the evidence reviewed.
Ruija offers interview audio and transcripts; access is application-based.4

Limit: No authorised Kven test dataset is established for this project. Ruija commercial evaluation and export rights need confirmation. Finnish or Meänkieli scores are not Kven scores.

Next: Agree access or commission consented recordings and human Kven transcripts, then test candidates.

Next: a newsroom evaluation.

Test suitability before committing to deployment.

  1. Agree the material

    Languages, programmes, audio conditions, rights and transcript conventions.

  2. Test consistently

    Freeze human references. Separate tuning from evaluation. Retain failures.

  3. Judge newsroom value

    Correction time, names, numbers, negation, omissions, latency and cost.

Deliverable: A recommendation per language—pilot, investigate further, or stop—with evidence and dependencies.