DavidErikMollberg/samromur_l2

Samrómur L2 22.09, with aligned text versions

140,099 recordings (149 hours) of 2,159 speakers whose native language is not Icelandic, reading sentences aloud in Icelandic. The largest groups are speakers of Polish, German and English, and there are 50 native languages in all. The recordings were crowd-sourced through samromur.is by the Language and Voice Lab at Reykjavík University with Almannarómur.

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. Every recording was also checked against its sentence by speech recognition, which matters here because the original release is mostly unverified.

Original release: Samrómur L2 22.09 on CLARIN-IS · Code to reproduce this dataset: Daviderikmollberg/icelandic-asr-data

from datasets import load_dataset
ds = load_dataset("DavidErikMollberg/samromur_l2", split="test", streaming=True)

Splits

split	recordings	hours	speakers
train	82,359	90.5	1,892
validation	26,663	28.4	193
test	27,704	27.3	187
validation_mismatch	1,740	1.6	143
test_mismatch	1,633	1.6	131

• The splits are the publisher's, with these changes:
  • Mismatches moved out of validation and test. Recordings that probably don't match their sentence go to validation_mismatch / test_mismatch, so the default evaluation sets are cleaner (see the speech-recognition check below).
  • 121 recordings with no split (all from one speaker) are in train (flag no_upstream_split).
• The splits share speakers.
  • 56 validation speakers and 40 test speakers are also in train.
  • 69 speakers are in both validation and test.
• Left out: 2,932 recordings (2.5 h).
  • 30 speakers read in another language instead of Icelandic, so all of their recordings are left out.
  • Recordings a person listened to and rejected are also left out.
• Age, gender and native language are as reported by the participants.

Checking the audio against the text

Only 4,957 recordings (all in train) were verified by the publisher. The rest, including all of validation and test, were not.

We transcribed every recording with our own Zipformer speech-recognition model (trained on other Icelandic corpora, not on this one) and compared the result with the sentence. meta_zip_cer is the character error rate and meta_zip_hyp is the transcript.

• The cut-off is a CER above 0.5 (flag asr_mismatch).
  • Of the recordings the publisher verified, only 0.3% are above it.
  • Of the unverified ones, 7.5% are above it.
• A person listened to 40 flagged validation/test recordings, 10 from each CER band:
  • above 0.7: all 20 were wrong (another sentence, silence, a cut-off recording, or another language)
  • 0.5–0.7: half were wrong
  • 0.3–0.5: mostly right; where the speaker misread, the text was corrected
• Validation and test recordings above 0.5 are in the *_mismatch splits.
• Train keeps them, flagged asr_mismatch (4,857 recordings). To filter them, drop rows whose flags contain asr_mismatch.

In the main validation and test splits, the median CER is 4.7%. Non-native speech is hard for the recognizer too, so a moderate CER is often accent, not a wrong recording.

Text columns

Example, 024687-1924858:

column	what it is	example
text_original	the prompt, verbatim from the release	Kíkir þú oft á Fótbolti.net?
text_written	the prompt as written Icelandic; speaker names in capitals written normally	Kíkir þú oft á Fótbolti.net?
text_written_plain	text_written without " „“ : ; / - ( )
text_cased	what was said, spoken form, names cased, spelled acronyms as letters	kíkir þú oft á Fótbolti punktur net
text_cased_joined	text_cased with spelled acronyms joined (H Í → HÍ)
text_normalized	text_cased in lower case	kíkir þú oft á fótbolti punktur net

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, Children and Queries):

• Spoken form. The text follows the prompt. The first word of a sentence is lower-cased unless it is a name, decided by corpus statistics, then the BÍN lexicon, then an LLM. Web addresses are read with "punktur".
• Acronyms and initials (160 recordings) were listened to.
  • MAI-Transcribe-2 and Azure is-IS (lexical form) transcribed each one, and an LLM drafted the spoken form.
  • 92 drafts follow the readings settled in Samrómur Children and Queries (flag reading_rule).
  • A person reviewed the other 68, nearly all in validation and test.
• Where a speaker misread, the text follows the audio, but only for the recordings a person listened to. In the rest, a misread word stays as the prompt; meta_zip_cer helps find these.
• 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, override, 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:
  • check against the audio: asr_mismatch
  • upstream state: upstream_unverified (never verified by the publisher), upstream_empty (the publisher's guess that the recording is empty), upstream_invalid, no_upstream_split
  • text: reading_rule, needs_listening / listen:<reason>, llm_casing, norm_diff, non_icelandic, foreign_words
• meta_zip_cer, meta_zip_hyp: our speech-recognition check
• meta_*: every column of the release's metadata.tsv, verbatim. These include meta_native_language, meta_is_valid and meta_marosijo_score (the publisher's audio–text match score, for 42,643 train recordings).

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

• Unverified audio. Apart from the 4,957 verified recordings and the 110 a person listened to, only the automatic check stands between a recording and its text. Some wrong recordings remain below the cut-off, and some correct but heavily accented ones are above it.
• Many prompts repeat, and validation and test share sentences with train.
• The splits are not speaker-independent (see Splits).

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; Mestrou, Thomas; Guðmundsson, Smári Freyr; Jónsson, Ólafur Helgi; Þorsteinsdóttir, Sunneva; Magnúsdóttir, Eydís Huld; Richter, Caitlin Laura; Pálsson, Ragnar; Gudnason, Jon, 2022, Samrómur L2 22.09, CLARIN-IS, http://hdl.handle.net/20.500.12537/263

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

@misc{mollberg2026samromurl2aligned,
  author       = {Mollberg, David Erik},
  title        = {Samr{\'o}mur L2 22.09 with aligned text versions and an audio--text check},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/DavidErikMollberg/samromur_l2}},
  note         = {Derived from Samromur L2 22.09 (CLARIN-IS, http://hdl.handle.net/20.500.12537/263). Code: \url{https://github.com/Daviderikmollberg/icelandic-asr-data}}
}