Samrómur: Crowd-sourcing large amounts of data

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 2311–2316 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 2311 Samrómur: Crowd-sourcing large amounts of data Staffan Hedström, David Erik Mollberg, Ragnheiður Þórhallsdóttir, Jón Guðnason Reykjavik University, Menntavegi 1, 102 Reykjavik Tiro, Laugavegur 163 Reykjavík {staffanh, ragnheidurth, jg}@ru.is {david.mollberg}@tiro.is Abstract This contribution describes the collection of a large and diverse corpus for speech recognition and similar tools using crowd-sourced donations. We have built a collection platform inspired by Mozilla Common Voice and specialized it to our needs. We discuss the importance of engaging the community and motivating it to contribute, in our case through competitions. Given the incentive and a platform to easily read in large amounts of utterances, we have observed four cases of speakers freely donating over 10 thousand utterances. We have also seen that women are keener to participate in these events throughout all age groups. Manually verifying a large corpus is a monumental task and we attempt to automatically verify parts of the data using tools like Marosijo and the Montreal Forced Aligner. The method proved helpful, especially for detecting invalid utterances and halving the work needed from crowd-sourced verification. Keywords: Speech corpora, Icelandic, Crowd Sourcing 1. Introduction The collection effort is part of the Icelandic national program for language technology (LT), a five-year pro- gram launched in October 2019 (Nikulásdóttir et al., 2020). One of the program’s goals was to collect a large and diverse collection of Icelandic speech data and make it readily available for use in automatic speech recognition (ASR) and similar fields. The initial goal was to collect 530 thousand utterances (Nikulásdóttir et al., 2017). That goal has been exceeded, and we have collected 1.5 million utterances that we estimate are about 2,250 hours of recorded speech. Samrómur is the largest prompted speech collection ef- fort for Icelandic so far and verifying the data is just as a monumental task as the collection itself. Therefore the data will be released in batches so that it can be- come of use as soon as possible. The first corpus to be released from the Samrómur collection was published on OpenSLR1 and it is also being added to the Linguis- tic Data Consortium (LDC). That subset contained 100 thousand utterances, or around 114 hours, of manually reviewed utterances for people aged 18 and up. The next pending subset from the Samrómur Collection will contain 137 thousand utterances, or 131 hours, of speech from children aged between 4-17. That will be the first Icelandic corpus intended for ASR with chil- dren’s speech. A third corpus containing 17 thousand utterances focused on queries from various sources will also be released on OpenSLR and LDC. As more and more data gets verified, additional coprora will but composed and released. Another recently released corpus that is part of the LT program is Talrómur (Sigurgeirsson et al., 2021). Al- though intended for text-to-speech (TTS), the corpus contains 213 hours of high-quality recordings …

Content truncated to 500 words