Whisper large-v3, fine-tuned for Finnish dialects

Speech-to-text for Finnish dialect speech: it hears dialect and writes standard Finnish with punctuation. A LoRA fine-tune of openai/whisper-large-v3 on dialect interviews, read sentences and 72 hours of Finnish parliament speech, so it also writes names, acronyms and terms more often right.

This is the main model since 21 Oct 2026: on the SKN dialect test 19.4 % of words wrong (the first published model, tag v1.0, 21.1 %; Whisper large-v3 40.4 %), level on read speech, and better on names and terms. The newest open model tested, NVIDIA Nemotron 3.5 Transcribe (released 28 Oct 2026), gets 47.6 % on the same test. The model published on 13 Oct 2026 stays available as revision="v1.0".

Trained at Haaga-Helia University of Applied Sciences in the GAIK project (Generative AI-enhanced Knowledge Management in Business), a research project led by Haaga-Helia with the University of Helsinki and Tampere University, co-funded by the European Union (ERDF).

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

📖 The story, with audio: suomeksi · in English · 🔊 listen and compare models · 🗂️ training data (SKN part)

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

Results

WER % after lower-casing and removing punctuation; the same decoding (transformers pipeline, greedy) for every model.

¹ This model was trained on other speakers and sessions of the same corpus, so these two tests are in its domain. Paired over segments against la-murre: names-and-terms set −1.35 points (95 % interval −2.04 to −0.80), SKN −0.35 (−0.73 to +0.04); FLEURS and LA-murre level. VoxPopuli moves by more than a point between two seeds of the same recipe, so it cannot separate these models. The names and terms were tagged by an LLM (Claude Opus 5.5) and the tags checked on a sample; the headline counts the three classes that check found reliable.

Use it

from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="GAIK-project/whisper-large-v3-fi-dialects",
    dtype="float16",
    device="cuda",
)
print(asr("piece.wav", generate_kwargs={"language": "fi", "task": "transcribe"})["text"])

Cut long audio into 15–25 s pieces at quiet points; the v1.0 card has code that does it.

Variants

• main: this model; revision="parl" pins it.
• v1.0: the first published model (13 Oct 2026), SKN + FLEURS + VoxPopuli: 21.1 / 7.6 / 10.8 %. Its card has the long-form, timestamp and judge measurements.
• la-murre: v1.0 + 56 h of LA-murre: 19.7 / 7.8 / 11.4 %.
• timestamps: v1.0's recipe with Whisper's timestamp tokens, for sentence times and long audio in one go.
• all-skn: v1.0's recipe on all 50 SKN municipalities.
• glossary: this model's recipe trained to take a word list (Sanasto: …) as a prompt; with the list it writes 93 % of names and terms right (this model 75 %).

How it was trained

1. 🎙️ Data. The la-murre variant's data: 15 042 SKN dialect segments, 15 529 LA-murre segments (56.5 h), FLEURS fi_fi train (2 704) and VoxPopuli fi train (7 504). Added: 37 740 segments (71.6 h) of the Aalto Finnish Parliament ASR Corpus 2015–2020 part 1, a seeded draw of 100 h leaving out the 2016-test speakers and the 2020 test sessions; segments with the corpus's <UNK> marker (an unknown word) were dropped.
2. ✍️ Training text. The parliament transcripts are lower case without punctuation. An LLM (Claude Haiku 5.5, and Sonnet 5.5 where Haiku changed a word) added punctuation and capitals with every word unchanged, checked by script; 362 segments neither could punctuate without changing a word were dropped. The other texts are as in the la-murre variant.
3. ⚙️ Fine-tuning. LoRA rank 64 on q_proj and v_proj, learning rate 1e-4, batch 8, two epochs (the best epoch was the second in every earlier run); validation loss 0.604 before, 0.232 after. Merged into the weights here.
4. 🖥️ Compute. 6 h 29 min on one AMD MI250X GPU on the LUMI supercomputer.

Where it is weaker

• VoxPopuli: 0.3 points more word errors than v1.0 (within the run-to-run noise of that set).
• Both dialect corpora are 1960s–70s interviews with elderly speakers; the parliament speech is formal. Everyday conversation, slang and Finnish-English mixing were not tested. One training run and one seed.

Credits and licences

• Developed by Akseli Leskinen (laguagu), Haaga-Helia University of Applied Sciences, in the GAIK project.
• Weights: Apache 2.0, as the base model openai/whisper-large-v3.
• Suomen kielen näytteitä (SKN), Kotus, via Kielipankki – The Language Bank of Finland, CC BY 4.0, urn:nbn:fi:lb-2020112937.
• Lauseopin arkiston murrekorpus (LA-murre), University of Turku and Kotus, via Kielipankki, CC BY-ND 4.0, urn:nbn:fi:lb-2020112935. In October 2026 Kielipankki (FIN-CLARIN) assessed that model weights trained on it most likely contain no copyright-protected part of the corpus. Its standard-Finnish rewrites are not published.
• Aalto Finnish Parliament ASR Corpus 2008–2020, version 2, Aalto University speech recognition group, from the plenary sessions of the Parliament of Finland, via Kielipankki, CLARIN PUB +BY +PRIV +ND, urn:nbn:fi:lb-2022052002. The model writes standard Finnish from the speech it is given and cannot reproduce the corpus. Neither the corpus nor its punctuated texts are published here, and nothing here is endorsed by the Parliament of Finland.
• FLEURS, Google, CC BY 4.0. VoxPopuli, Meta, CC0.
• Computing on LUMI. We acknowledge CSC – IT Center for Science, Finland for awarding this project access to the LUMI supercomputer, owned by the EuroHPC Joint Undertaking, hosted by CSC (Finland) and the LUMI consortium through CSC – IT Center for Science, Finland.
• Made in the GAIK project, co-funded by the European Union (ERDF).