LaboroTVSpeech-ASV, also shortened as Laboro-ASV, is a dataset for Japanese speaker verification. Similar to LaboroTVSpeech dataset, the source data comes from recordings of Japanese terrestrial TV programs.
LaboroTVSpeech serves as an excellent dataset for tasks including Japanese speech recognition, while LaboroTVSpeech-ASV focuses on Japanese speaker verification.
Comparing to large-scale speaker verification datasets like VoxCeleb and CN-Celeb, Laboro-ASV has fewer POIs and total hours, but the average number of utterances per POI is significantly larger.
- EER is adopted as the evaluation metric, and lower is better.
- Laboro-ASV dataset works as an effective supplemental dataset for Japanese speaker verification.
- The subset Laboro-ASV-50 includes only the top 50 POIs ranked by the number of utterances, and it also gives the best performance among all tested datasets when used as a supplemental dataset.
- This research has been presented at ASJ 2023 Autumn Meeting.
- For more details of this research, please refer to our paper (the link will be added soon) and our poster.
- Please cite this repository if the Laboro-ASV dataset has been used in your research:
Laboro.AI Inc., Laboro-ASV, https://github.com/laboroai/LaboroTVSpeech-ASV, 2023
The LaboroTVSpeech-ASV dataset is published in the same way as LaboroTVSpeech. To obtain the dataset, please
(1) refer to the "LaboroTVSpeechのご利用について" section in this page
or
(2) send an inquiry directly to our company via this page
The content of this repository (excluding LaboroTVSpeech-ASV dataset itself) is released under Apache License v2.