tiro-is/kennsluromur

KENNSLURÓMUR - ICELANDIC LECTURES

[Icelandic]

Kennslurómur - Íslenskir fyrirlestrar er safn af hljóðskrám og samsvarandi texta úr kennslufyrirlestrum sem teknir voru upp í áföngum í Háskólanum í Reykjavík og Háskóla Íslands. Þetta safn má nota við þjálfun talgreina.

Fyrirlesararnir gáfu upptökurnar sínar sem síðan voru talgreindar með talgreini, næst var frálagið lesið og leiðrétt af hópi sumarnema og að lokum var allur texti yfirfarinn af prófarkalesara.

Í þessu safni eru 51 klukkustund af hljóðskrám sem dreifast á 171 fyrirlestur frá 11 fyrirlesurum.

[English]

Kennslurómur - Icelandic Lectures is a collection of audio recordings and their corresponding segmented transcripts from class lectures recorded at Reykjavik University and the University of Iceland. This material was compiled for the training of speech recognition models.

The lectures were donated by each lecturer, then transcribed with an Icelandic speech recognizer, then manually corrected by human transcribers and finally verified by a proofreader.

This release contains 51 hours divided between 171 lectures from 11 lecturers.

LECTURE TOPICS

The topic of the lextures cover a diverse range of university level subjects.

Linguistics         	   15 lectures	1 speaker	7,12 hours
Computer science	       33 lectures	3 speakers	15,3 hours
Labour market economics	   13 lectures	1 speaker	1,91 hours
Engineering                64 lectures	3 speakers	11,3 hours
Legal studies	           25 lectures	2 speakers	7,52 hours
Business intelligence	    1 lecture	1 speaker	19,2 minutes
Psychology	               10 lectures	1 speaker	3,03 hours
Sports science             10 lectures	1 speaker	4,79 hours

STRUCTURE

SPEAKERS.tsv                        -  Lists the speakers (lecturers) and their IDs.
LECTURES.tsv                        -  Lists all lectures. See header for the format.
DOCS/
    transcription_guidelines_is.txt -  Transcription guidelines in Icelandic.
LICENSE.txt                         -  Description of the license.
prerp_for_training.py               -  An example data preparation script for KALDI.
<SPK-ID>/                           -  A directory per speaker.
    <LECTURE-ID>.wav                -  Audio recording of the entire lecture.
    <LECTURE-ID>.txt                -  Transcript of the entire lecture in 1 to 
                                       40 second segments. Tab separated list with the
                                       fields: segment ID, start time in milliseconds, 
                                       end time in milliseconds and utterance text.

Alignment and segmentation

The segments are mostly split on sentence boundaries. Each segment ranges from a few seconds to roughly 40 seconds in duration. The recordings and transcripts were automatically aligned using either Montreal Forced Aligner or the aligner Gentle. The alignment quality was tested by training an acoustic model in Kaldi and rejected segments due to alignment issues. Recordings with an abnormally high number of faulty segments were manually aligned. This means that there are likely still some imperfectly aligned segments, but due to resource constraints, they were not manually checked and verified.

Training, development and testing sets

Every segment has been marked as either train, dev or eval. This can be seen in the <SPK-ID>/<LECTURE-ID>.txt files. There are a few speakers in this dataset creating training sets without overlap of speakers is not possible without holding out a large portion of the data. Therefore, it was decided to randomly assign each speaker's segments proportionally 80/10/10 (train, dev, eval) based on the duration of each segment.

FORMAT

Sampling rate     16000 Hz
Audio format      16 bit PCM RIFF WAVE
Language          Icelandic
Type of speech    Single speaker spontaneous and scripted speech with minimal
                  backspeech.
Media type        Recorded university lectures, a mixture of prerecorded 
                  classes and in-class recordings.

SPECIAL ANNOTATIONS

Three types of special annotations are found the transcripts:

[UNK]           Unintelligible, spoken background noise

[HIK: <stubs>]  Hesitation, where <stubs> can be a comma separated list
                of false start (often partial) words.

[<IPA sym>]     Standalone IPA phones are transcribed in brackets which
                only appear in "Icelandic linguistics" lectures.
                E.g. "Þannig fáum við eins og raddað b, [p] [p] [p] 
                „bera bera“.".

LICENSE

The audio recordings (.wav files) are attributed to the corresponding lecturer in the file SPEAKERS.tsv. Everything else is attributed to Tiro ehf.

Published with a CC BY 4.0 license. You are free to copy and redistribute the material in any medium or format, remix, transform and build upon the material for any purpose, even commercially under the following terms: You must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your use.

Link to the license: https://creativecommons.org/licenses/by/4.0/

ACKNOWLEDGMENTS

This project was funded by the Language Technology Programme for Icelandic 2019-2023. The programme, which is managed and coordinated by Almannarómur, is funded by the Icelandic Ministry of Education, Science and Culture.