Digisensus/lithuanian-dialect-asr-eval

Lithuanian dialect ASR evaluation bundle

Everything needed to check or extend the evaluation of the study Lithuanian speech recognition: effects of dialect training and transcript spelling: the test sets, every reference layer, the hypotheses of every evaluated model on every test set, and all scores. Code and scorer: github.com/Digisensus/lithuanian-dialect-asr.

Test sets

Config	Test set	Clips	Hours	Speakers	References
dial_test	LIEPA-3 dialect test: 12 unseen speakers, 3 per region	5,607	5.01	12	dialect spelling (dial), standard spelling (std), word alignment
dial_seen	LIEPA-3 dialect, held-out clips of 12 training speakers	1,144	1.01	12	as dial_test
phone_test	LIEPA-3 telephone test (not speaker-disjoint)	4,206	4.36	–	spoken form (text)
fleurs_test	FLEURS lt_lt test	986	2.97	–	raw transcript (text)
cv_test	Common Voice 19.0 lt test	4,925	7.15	260	raw transcript (text)

test_sets/<test set>.txt lists the clip ids; references/<test set>.jsonl holds the reference text. Audio is not included: the LIEPA-3 clips are in lithuanian-dialect-speech-liepa-3-100h-punctuated and lithuanian-phone-speech-liepa-3-429h-punctuated (same ids), FLEURS in google/fleurs and Common Voice 19.0 in its CC0 mirror fsicoli/common_voice_19_0. Both benchmarks are pinned: FLEURS lt_lt test at revision 168de341b3db6859a9bac1c50a2ef5e3b47647e0 (refs/convert/parquet), clip fleurs-lt-NNNN = row NNNN; Common Voice lt test at revision 590c8abec6cf7c8d06e650f1438e60332a796e11, clip cv19-lt-N = file common_voice_lt_N.mp3, speaker = hashed client id.

The dialect references follow the alignment format of the standard-spelling layer: dial is the dialect spelling, std the standard spelling, and alignment pairs every dialect word with its standard form, so a hypothesis word can be accepted in either form (either-form WER). The standard-spelling references are semi-automatic (see the dataset card).

Hypotheses

hypotheses/<system>/<test set>.jsonl, one {"id", "hyp"} per clip: greedy decoding, each model's own settings, every system on every test set (105 files).

System	Run
B_parakeet-tdt-0.6b-v3	zero-shot baseline nvidia/parakeet-tdt-0.6b-v3
B_whisper-large-v3	zero-shot baseline openai/whisper-large-v3
D80.dial_s1_a2	ltd26/D80.dial/s1/a2, recipe D80.dial (model)
D80.std_s1_a1	ltd26/D80.std/s1/a1, recipe D80.std (model)
S+D12.dial_s1_a1	ltd26/S+D12.dial/s1/a1, recipe S+D12.dial
S+D25.dial_s1_a1	ltd26/S+D25.dial/s1/a1, recipe S+D25.dial
S+D25.std_s1_a1	ltd26/S+D25.std/s1/a1, recipe S+D25.std
S+D50.dial_s1_a1	ltd26/S+D50.dial/s1/a1, recipe S+D50.dial
S+D50.std_s1_a1	ltd26/S+D50.std/s1/a1, recipe S+D50.std
S+D80-A.std_s1_a1	ltd26/S+D80-A.std/s1/a1, recipe S+D80-A.std
S+D80-D.std_s1_a1	ltd26/S+D80-D.std/s1/a1, recipe S+D80-D.std
S+D80-S.std_s1_a2	ltd26/S+D80-S.std/s1/a2, recipe S+D80-S.std
S+D80-Z.std_s1_a2	ltd26/S+D80-Z.std/s1/a2, recipe S+D80-Z.std
S+D80.dial_s1_a3	ltd26/S+D80.dial/s1/a3, recipe S+D80.dial (model)
S+D80.dial_s2_a1	ltd26/S+D80.dial/s2/a1, recipe S+D80.dial
S+D80.std_s1_a1	ltd26/S+D80.std/s1/a1, recipe S+D80.std (model)
S+D80.std_s2_a1	ltd26/S+D80.std/s2/a1, recipe S+D80.std
S+X80_s1_a2	ltd26/S+X80/s1/a2, recipe S+X80 (model)
S+X80_s2_a1	ltd26/S+X80/s2/a1, recipe S+X80
S_s1_a2	ltd26/S/s1/a2, recipe S (model)
S_s2_a1	ltd26/S/s2/a1, recipe S

Scores

scores.csv: one row per system × test set × metric, produced by score.py from the stored hypotheses (columns eval_id, layer_id, metric, value, n_words, scorer_version). Metrics: WER against dialect references (wer_dialect), standard references (wer_standard) and either form (wer_either), CER, mean WER over speakers, per-region WER (dialect test) and either-form errors by change tag. Normaliser: lowercase, NFC, punctuation to space.

Control pool

control_pool/ documents the study's control condition: 89,892 clips of other LIEPA-3 spontaneous speech (radio, dictaphone, TV; not telephone, not dialect), matched to the dialect training set by gender, age group and clip length. xspon.train.txt lists the clip ids, xspon.train.meta.tsv their LIEPA-3 path, duration, gender, age group and source, and build_xpool.py rebuilds the list from the original LIEPA-3 release.

Licence and attribution

The references, hypotheses, scores and lists are © 2026 Digisensus, licensed under CC BY 4.0. The LIEPA-3 transcripts come from LIEPA-3 by Vytautas Magnus University, Vilnius University and the Institute of the Lithuanian Language (CC BY 4.0); the FLEURS transcripts from Google (CC BY 4.0); the Common Voice transcripts from Mozilla Common Voice (CC0).

Citation

@misc{digisensus_lithuanian_dialect_asr_eval_2026,
  title        = {Lithuanian dialect ASR evaluation bundle},
  author       = {{Digisensus}},
  year         = {2026},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/datasets/Digisensus/lithuanian-dialect-asr-eval}
}