DavidErikMollberg/samromur_queries

Samrómur Queries 21.12, with aligned text versions

17,474 validated recordings (20 hours) of 3,808 speakers reading questions aloud in Icelandic. The recordings were crowd-sourced through samromur.is by the Language and Voice Lab at Reykjavík University with Almannarómur. The questions are the kind a voice assistant gets, such as opening hours, trivia and word meanings. There are also short lines from plays.

This copy keeps the original audio: 16 kHz FLAC, unchanged, embedded in Parquet so it streams. It also adds several aligned versions of the text, so the same data can train models that output spoken-form text, cased text, or written text with punctuation.

Original release: Samrómur Queries 21.12 on CLARIN-IS · Code to reproduce this dataset: Daviderikmollberg/icelandic-asr-data

from datasets import load_dataset
ds = load_dataset("DavidErikMollberg/samromur_queries", split="test", streaming=True)

Splits

split	recordings	hours	speakers
train	14,139	16.0	2,704
validation	1,728	2.0	534
test	1,607	2.0	570

• The splits are the publisher's. Validation and test are gender-balanced, and no speaker in either is in train.
• One train recording is left out (013380-0406805): its file is undecodable.
• Age is as reported by the participants. It is mostly a range (40-49); children gave a year (10).

Text columns

Example, 010724-0286368:

column	what it is	example
text_original	the prompt, verbatim from the release	Friðfinnur, hvað er opið lengi í HR?
text_written	the prompt as written Icelandic; speaker names in capitals (LÁRUS:) written normally	Friðfinnur, hvað er opið lengi í HR?
text_written_plain	text_written without " „“ : ; / - ( )	Friðfinnur, hvað er opið lengi í HR?
text_cased	what was said, spoken form, names cased, spelled acronyms as letters	Friðfinnur hvað er opið lengi í H R
text_cased_joined	text_cased with spelled acronyms joined (H R → HR)	Friðfinnur hvað er opið lengi í HR
text_normalized	text_cased in lower case	friðfinnur hvað er opið lengi í h r

text_normalized is always exactly text_cased.lower(). Spoken forms contain only letters and spaces.

How the text was made (the same recipe as our Samrómur 21.05 and Samrómur Children):

• Casing. The first word of a sentence is lower-cased unless it is a name. This is decided by corpus statistics first, then the BÍN lexicon, then an LLM (flag llm_casing).
• Speaker names in capitals (LÁRUS:) are written as normal words.
• Readings we can't know from the text were listened to. The prompts have no digits, so these are acronyms and one abbreviation, 72 recordings.
  • MAI-Transcribe-2 and Azure is-IS (lexical form) transcribed each one, and an LLM drafted the spoken form.
  • A person confirmed one reading per acronym:
    • spelled out (H Í): HÍ, HR, ÁTVR, EM, HM, IMDB, NBA, ISBN, VSOP
    • said as words, kept in capitals: IKEA, NOCCO, DAS, SETI
  • 45 train drafts follow those readings (flag reading_rule).
  • A person reviewed the other 27, including every one in validation and test.
• Where the speaker read something else, the text follows the audio. One speaker read ÁTVR as Á tvar.
• Typos. Three prompts had typos, which are fixed in every version except text_original (flag prompt_fixed): ‚ser → sér, and a ? inside a word (máli?lýsa).
• Written form. The prompt, cleaned. Where the spoken form needed words in place of symbols, an LLM wrote the written form, accepted only if every other word stayed the same.

text_cased_src and text_written_src record where each row's text came from (rule, rule+bin, rule+llm, llm+asr, review, original, llm).

Other columns

• id (<speaker>-<utterance>, as in the audio file names), speaker_id, gender (female, male, other, unknown; self-reported), age, duration, split, corpus
• flags: llm_casing, reading_rule, needs_listening / listen:<reason>, prompt_fixed, norm_diff, foreign_words
• meta_*: every column of the release's metadata.tsv, verbatim (meta_sentence, meta_sentence_norm, meta_native_language, meta_dialect, meta_created_at, meta_is_valid, meta_user_agent, meta_status, …)

For k2/icefall with Lhotse, the companion code exports to Lhotse Shar with all text versions in supervision.custom: https://github.com/Daviderikmollberg/icelandic-asr-data

Known limitations

• Many prompts repeat: 2,211 distinct questions are read more than once, often with only the name at the start changed (<name>, hvað er opið lengi í HR?).
• Train is mostly women: 9,621 of 14,139 recordings.
• The publisher validated each recording (meta_is_valid). We listened only to the 72 recordings with ambiguous readings.

Licence and citation

CC BY 4.0, the same as the original release. Please cite the original corpus:

Hedström, Staffan; Fong, Judy Y.; Þórhallsdóttir, Ragnheiður; Mollberg, David Erik; Guðmundsson, Smári Freyr; Jónsson, Ólafur Helgi; Þorsteinsdóttir, Sunneva; Magnúsdóttir, Eydís Huld; Gudnason, Jon, 2021, Samromur Queries 21.12, CLARIN-IS, http://hdl.handle.net/20.500.12537/180

If you use the added text versions (text_written, text_written_plain, text_cased, text_cased_joined, text_normalized), the reading decisions or the flags, please also cite this version:

@misc{mollberg2026samromurqueriesaligned,
  author       = {Mollberg, David Erik},
  title        = {Samr{\'o}mur Queries 21.12 with aligned text versions: original, written, spoken-form cased and normalized transcripts},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/DavidErikMollberg/samromur_queries}},
  note         = {Derived from Samromur Queries 21.12 (CLARIN-IS, http://hdl.handle.net/20.500.12537/180). Code: \url{https://github.com/Daviderikmollberg/icelandic-asr-data}}
}