SKN Finnish dialects, cut for speech recognition

Finnish dialect interviews from Suomen kielen näytteitä (SKN), cut into 2–25 s segments with a standard-Finnish transcript, split by municipality for training and testing speech recognition. Made at Haaga-Helia University of Applied Sciences in the GAIK project (led by Haaga-Helia with the University of Helsinki and Tampere University, co-funded by the European Union, ERDF) to train GAIK-project/whisper-large-v3-fi-dialects.

GAIK project · Haaga-Helia University of Applied Sciences · developed by Akseli Leskinen (laguagu)

🤗 The model trained on it (dialect test word error rate 40.4 % → 21.1 %) · 📖 The story, with audio: suomeksi · in English · 🔊 listen and compare models

📊 scripts, method and all results: GAIK-project/finnish-asr · 🧰 gaik-toolkit · 💻 GAIK on GitHub · 🌐 gaik.ai

The test municipalities are SKN numbers 5, 10 … 50: every fifth booklet of the series, in publication order. By Kielipankki's dialect classification they cover the south-western, south-western transitional, Häme, Savo and south-eastern dialects; no Ostrobothnian or Peräpohjola municipality is in the test (all nine are in train). No test interviewee is in training. Interviewer speech is kept, and SKN's fieldworkers recorded several municipalities, so a few test segments may contain an interviewer's voice that is also heard in training.

Fields

How it was made

1. SKN's sentence annotations of all speakers, in time order, are joined while the gap is under 1.5 s and the segment stays under 25 s; segments under 2 s or with fewer than 3 words are dropped.
2. Test and validation keep a reproducible sample (seed 0) of 60 and 100 segments per municipality.
3. For train and validation, Claude Opus 5.5 added punctuation and capitals to the standard-Finnish text, 20 segments per call, text only. An answer was kept only if its words were exactly the human transcript's; 15 024 of 15 042 train segments passed, the others have text_source = human. Test transcripts are untouched human text.

The scripts that made it are public in GAIK-project/finnish-asr, which also holds the method and all results (summary in gaik-toolkit · finnish_dialect_asr):

• prepare_skn.py cuts SKN into these segments and splits them by municipality;
• prepare_targets.py adds the punctuation and keeps an answer only if its words are unchanged;
• prepare_timestamps.py turns SKN's sentence annotation into sentence times inside every segment, results/labels/skn_parts.jsonl (by id: each sentence's start, end and word count), and into ten windows of about 4 minutes, one per test municipality, for long-form tests, results/labels/skn_longform.jsonl. The timestamps revision of the model was trained with the sentence times of the training segments.

Caveats

• text_dialect was empty in every row of the first upload (13 Oct 2026); it was filled on 14 Oct 2026. The audio, the other fields and the splits did not change.
• 1960s–70s interviews with mostly elderly speakers, recorded in varying conditions.
• Interviewer speech is included where it falls inside a segment, as in SKN's annotation.
• Punctuation in train and validation is model-written, checked only for unchanged words.
• In a few dozen segments SKN's normalised transcript itself keeps some punctuation or dialect notation; text_normalized (and so text) leaves it as SKN has it.

Source and licence

Developed by Akseli Leskinen (laguagu), Haaga-Helia University of Applied Sciences, in the GAIK project.

Suomen kielen näytteitä, Kotimaisten kielten keskus (Kotus), via Kielipankki – The Language Bank of Finland, urn:nbn:fi:lb-2020112937, CC BY 4.0. Original corpus: kielipankki.fi/corpora/skn, download kielipankki.fi/download/SKN. This derived dataset is also CC BY 4.0; credit Kotus and Kielipankki as the source.