ERISLab/LisTAya-transcripts

LisTAya transcripts: the test-set evaluations of the LisTAya study

This dataset holds the reference and the model output for every utterance of every test-set evaluation in the paper Does Regional Decoder Specialization Help Low-Resource ASR Based on the SLAM-ASR Framework? (ROCLING 2026). The trained models are described in the model card ERISLab/LisTAya and listed in the collection https://huggingface.co/collections/ERISLab/listaya-slam-asr-projectors-on-tiny-aya-rocling-2026-6ab031414095f3df8805ed7b.

Configs

Each config has one split, test, with one row per utterance per run. A run is one model evaluated on one test set.

Config	Models	Runs	Utterances
listaya	The trained LisTAya checkpoints: the four Tiny Aya regional decoder variants (Global, Earth, Fire, Water) for each of the twelve languages, each checkpoint evaluated on the test sets of its own language.	172	933,788
baselines	Three off-the-shelf models: Qwen/Qwen2-Audio-7B, mistralai/Voxtral-Mini-3B-2507, openai/whisper-medium.	127	662,017

Test sets

There are 43 test sets, all test splits:

• out-of-domain: the FLEURS test split (google/fleurs) of 11 languages; FLEURS does not cover Kreol Seselwa.
• in-domain: the WorldSpeech test split of the variety each language trains on, one per language (12). Hausa's is ha_td, because its checkpoints were selected on the ha_ng test split. Kreol Seselwa's is the test_clean split of ERISLab/WorldSpeech config crs_sc; every other language's comes from disco-eth/WorldSpeech.
• trained variety: the WorldSpeech test split of a second variety in the training data: sw_tz, ur_in (2).
• held-out variety: the WorldSpeech test splits of country varieties absent from training (18): en_au, en_jm, en_ke, en_nz, en_pk, en_sl, en_zm, es_ar, es_cl, es_co, es_es, es_pe, es_pr, es_py, es_uy, fr_cd, fr_ci, ta_lk.

Every baseline is evaluated on all 43 test sets, except that mistralai/Voxtral-Mini-3B-2507 and openai/whisper-medium produced no output on Kreol Seselwa, so baselines holds 127 runs (3 x 43 - 2).

Columns

Column	Type	Content
model	string	Hub id of the evaluated model. In listaya it is the checkpoint repository, whose name gives the decoder, the training language and the checkpoint step.
decoder	string	listaya only: the Tiny Aya decoder variant, Global, Earth, Fire or Water.
language	string	Language of the test set: Amharic, English, French, Hausa, Hindi, Indonesian, Kreol Seselwa, Marathi, Spanish, Swahili, Tamil, Urdu.
language_code	string	ISO 639 code of language.
eval_dataset	string	Hub path of the test set.
eval_config	string	Config of the test set: the FLEURS language or the WorldSpeech country variety, such as en_au.
eval_split	string	Split of the test set.
domain	string	in-domain, out-of-domain, trained variety or held-out variety, as defined above.
sample_index	int64	Position of the utterance among the evaluated utterances of the split, in split order.
duration	float64	Length of the audio in seconds.
reference	string	The normalised reference transcript.
hypothesis	string	The normalised model output.

Text normalisation

The references are the transcription column of FLEURS and the human_transcript column of WorldSpeech. References and model outputs pass through the same normaliser before scoring, for every language: lowercasing; removal of text inside square brackets, angle brackets and parentheses; Unicode NFKD decomposition, with nonspacing combining marks (category Mn) deleted and all other marks, symbols and punctuation replaced by a space; removal of any remaining character that is neither a word character nor whitespace; and collapsing of whitespace. In scripts that write vowels as combining signs, such as Devanagari and Tamil, the normalised text therefore keeps only part of each syllable. Utterances whose normalised reference is empty are left out of the evaluation, so sample_index equals the row index of the source split only where none were left out.

Relation to the paper

The CER of a run is the corpus-level character error rate of its rows in percent, computed with the cer metric of the evaluate library and rounded to two decimals. Recomputed from these rows, it equals the CER the paper uses for all 299 runs. The paper's per-test-set tables print, for each test set, the lowest of these CERs among the three baselines and among the language's four trained checkpoints: the in-domain and out-of-domain test sets in the table that compares training with the baselines, and the trained and held-out varieties in the table of country varieties.

Load the transcripts and score a run

import evaluate
from datasets import load_dataset

rows = load_dataset("ERISLab/LisTAya-transcripts", "listaya", split="test")
run = rows.filter(lambda r: r["model"] == "ERISLab/q2a_openai_whisper-medium_CohereLabs_tiny-aya-global_ws-en_us-500" and r["eval_dataset"] == "google/fleurs")
cer = evaluate.load("cer").compute(references=run["reference"], predictions=run["hypothesis"])
print(f"{100 * cer:.2f}")  # 6.22

Licence and attribution

The dataset is released under CC BY-NC 4.0. The references are normalised transcripts from FLEURS (google/fleurs, CC BY 4.0) and WorldSpeech (disco-eth/WorldSpeech, CC BY-NC 4.0), whose non-commercial term this release follows. The Kreol Seselwa references come from ERISLab/WorldSpeech, config crs_sc, recorded sessions of the National Assembly of Seychelles.

Citation

@inproceedings{rios-etal-2026-regional,
    title = "Does Regional Decoder Specialization Help Low-Resource {ASR} Based on the {SLAM}-{ASR} Framework?",
    author = "Rios, Edwin Arkel  and
      Zaruma, Jocelyn  and
      Ewoorkar, Girish  and
      Sourabh, Sneh  and
      Mack, Julian  and
      Juan, Hung-Hui  and
      Huang, Stephen  and
      Lai, Bo-Cheng",
    booktitle = "Proceedings of the 38th Conference on Computational Linguistics and Speech Processing (ROCLING 2026)",
    year = "2026",
    publisher = "Association for Computational Linguistics"
}