<< List of all deliverables

D1.2.1: Updated backend of existing ASRs

Project: FIN-CLARIAH
Grant agreement: Research Council of Finland no. 367751
Start date: 01-01-2026
Duration: 24 months

WP 1.2: Report on Speech processing and annotation
Date of reporting: 19-10-2026

Report author: Sam Hardwick (CSC)
Contributors: Mikko Kurimo, Yaroslav Getman (Aalto University)
Deliverable location: https://github.com/CSCfi/kielipankki-rahti/tree/KP-12001-wav2vec currently, https://github.com/CSCfi/kielipankki-rahti when merged

Keywords for the deliverable page: ASR; automatic speech recognition; wav2vec; Finnish; Finland Swedish; Sámi

Description

The previously existing ASR-related backends hosted at https://kielipankki.rahtiapp.fi/audio/asr and https://kielipankki.rahtiapp.fi/audio/forced-align were

1. Finnish-only
2. Lacking diarisation
3. Lacking text postprocessing (for punctuation and capitalisation
4. Based on kaldi-serve, for which new models were no longer being made and was difficult to maintain

This deliverable improves all of these aspects. It brings:

1. A new and improved Finnish model, and additionally a Finland Swedish and Sámi model (see Getman et al. 2025)
2. A diarisation model based on pyannotate
3. A reasonably good postprocessing model based on xlm-roberta for Finnish and Swedish, and a rudimentary one for Sámi
4. Based on wav2vec, a modern and well-supported platform

The new endpoints are available at

https://kielipankki-dev.rahtiapp.fi/audio/asr/{lang}/submit_file

with /fi/, /sv-FI/ and /sme/. The endpoints accept URL parameters diarize=false and punctuate=false to suppress those stages, and num_speakers=N to fix the diarisation speaker count.

Frontend support in the form of tekstiks.ee support and a command-line script are forthcoming.

FIN-CLARIAH project has received funding from the European Union – NextGenerationEU instrument and is funded by the Research Council of Finland under grant number 367751.

Last modified on 2026-11-06