-------------------------------------------------------------------------------- Samrómur L2 22.09 -------------------------------------------------------------------------------- Language : Icelandic Authors : Staffan Hedström, Judy Y. Fong, Ragnheiður Þórhallsdóttir, David Erik Mollberg, Thomas Mestrou, Smári Freyr Guðmundsson, Ólafur Helgi Jónsson, Sunneva Þorsteinsdóttir, Eydís Huld Magnúsdóttir, Caitlin Laura Richter, Ragnar Pálsson, Jon Gudnason Recommended use : speech recognition, speaker verification, speaker identification and speaker enrollment -------------------------------------------------------------------------------- Description -------------------------------------------------------------------------------- This release of data from the Samrómur collection focuses on speakers where Icelandic is not their native language. The corpus contains 143,031 (151.8 hours) of mostly un-verified speech recordings in Icelandic. The corpus is a result of the crowd-sourcing effort run by the Language and Voice Lab (LVL) at Reykjavik University, in cooperation with Almannarómur, the Icelandic Center for Language Technology. The recording process has started in October 2019 and ended September 2022. The present edition of the corpus has been authorized for release in September 2022. The aim is to create an open-source speech corpus to enable research and development for Icelandic Language Technology. The corpus consists of audio recordings and a metadata file containing the prompts read by the participants. To see more open resources developed by the Language and Voice Lab (LVL) see the GitHub repository at https://github.com/cadia-lvl/samromur-asr -------------------------------------------------------------------------------- Corpus Characteristics -------------------------------------------------------------------------------- - Only a small part of the corpus is validated. The corpus contains 4,957 validated utterances and the rest (138,074) of the utterances are not validated. - The utterances were recorded by a smartphone or the web app. - Participants self-reported their age group, gender and native language. - Participants are from 6 to 80+ years. - The corpus contains 143,031 utterances from 2,189 speakers, totalling 151.8 hours. - The amount of data from female speakers is 101h28m and the amount of data from male speakers are 46h4m and the amount of data from speakers with an unknown gender information is 4h16m. - The number of female speakers is 761, and the number of male speakers is 473. The number of speakers with unknown gender information is 955. - The number of utterances from female speakers is 97,075; the utterances from male speakers are 42,762; and the utterances from speakers with unknown gender information is 3,194. - The corpus is split into train, dev, and test sets. Lengths of the sets are: train = 91.4h, test = 29.3h, dev = 31.1h. For more such subsets please use Samrómur 21.05, Samrómur Queries 21.12 or Samrómur Children 21.09. - If any of the information in the metadata is unavailable this will be indicated with a NAN in the metadata file. -------------------------------------------------------------------------------- Collection Procedure -------------------------------------------------------------------------------- The data was collected using the website https://samromur.is, the code of which is available at https://github.com/cadia-lvl/samromur. The collection procedure is well described in "Samrómur: Crowd-sourcing Data Collection for Icelandic Speech Recognition" [1] and "Samrómur: Crowd-sourcing large amount of data” [3]. The original audio was collected at 44.1 kHz or 48 kHz sampling rate as _.wav files, which was down-sampled to 16 kHz and converted to _.flac. Each recording contains one read prompt from a script. The script contains 89,083 unique prompts, 956,989 tokens and 48,875 word types. Each time a device visits the website for the first time they are assigned a client id, this client id together with a combination of gender, age and native language was used to assign the speaker id. If any of these variables were changed, a new speaker id was also created. The corpus is distributed with a metadata file with detailed information on each utterance and speaker. The metadata file is encoded as UTF-8 Unicode. The prompts were gathered from a variety of sources, mainly from The Icelandic Gigaword Corpus, which is available at http://clarin.is/en/resources/gigaword. The corpus includes prompts found in novels, news, and plays. The prompts also came from the Icelandic Web of Science (https://www.visindavefur.is/). Some prompts were donated from Icelandic course material. Prompts were pulled from these sources if they met the criteria of having only letters which are present in the Icelandic alphabet, and if they are listed in DIM: Database Icelandic Morphology [2]. Finally, there are also synthesized prompts consisting of a name followed by a question, in order to simulate dialogue with a smart device. -------------------------------------------------------------------------------- Data Format Specifics -------------------------------------------------------------------------------- - Text : The corpus does not contain separate transcription or prompt files. The metadata file contains the prompts in their original text form, as the participants saw them, and also in their normalised form. - Audio: The distributed audio files are encoded at 16 kHz sampling rate, 16 bit linear PCM, 1 channel, \*.flac format. The corpus is split into train, dev and test subsets. Each subset contains folders that correspond to speaker IDs and the audio files inside use the following naming convention: {speaker_ID}-{utterance_ID}.flac. -------------------------------------------------------------------------------- Citation -------------------------------------------------------------------------------- When publishing results based on the corpus please refer to: Hedström et al. "Samrómur L2 22.09". Web Download. Reykjavik University: Language and Voice Lab, 2022. Contact: Jon Gudnason (jg@ru.is) License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/legalcode) -------------------------------------------------------------------------------- Acknowledgements -------------------------------------------------------------------------------- This project was funded by the Language Technology Programme for Icelandic 2019-2022. The programme, which is managed and coordinated by Almannarómur, is funded by the Icelandic Ministry of Education, Science and Culture. Special thanks to the assisting LVL members and summer students for all the hard work. -------------------------------------------------------------------------------- Stats for the dataset -------------------------------------------------------------------------------- Age and gender split: | | Total | Test | Dev | Train | | ---------------- | ----- | ----- | ----- | ----- | | 0-19: | 58.4% | 46.6% | 57.7% | 63.2% | | 20-39: | 14.5% | 21.9% | 10.9% | 13.1% | | 40-59: | 24.4% | 31.3% | 29.2% | 20.2% | | 60-79: | 1.1% | 0.0% | 0.9% | 1.5% | | 80+: | 0.6% | 1.2% | 1.2% | 0.3% | | ---------------- | ----- | ----- | ----- | ----- | | Female: | 67.9% | 66.3% | 65.8% | 69.3% | | Male: | 29.9% | 32.9% | 32.0% | 28.0% | | Other: | 2.2% | 0.8% | 2.1% | 2.8% | | ---------------- | ----- | ----- | ----- | ----- | | Duration (h): | 151.8 | 29.3 | 31.1 | 91.4 | | Unique speakers: | 2189 | 203 | 220 | 1906 | Amount of utterances in each subset: train: 83,404 dev: 29,748 test: 29,758 Total speakers and utterances: Speakers: 2,189 Utterances: 143,031 Average utterance length: 3.82s -------------------------------------------------------------------------------- References -------------------------------------------------------------------------------- [1] Mollberg et al. "Samrómur: Crowd-sourcing Data Collection for Icelandic Speech Recognition," 12th International Conference on Language Resources and Evaluation (LREC), France, 2020. [2] Bjarnadóttir et al. "DIM: The Database of Icelandic Morphology". Proceedings of the 22nd Nordic Conference on Computational Linguistics (NoDaLiDa), Finland. 2019. [3] Hedström et al. "Samrómur: Crowd-sourcing large amount of data”, 13th International Conference on Language Resources and Evaluation (LREC), France, 2022. --------------------------------------------------------------------------------