xtts split appears to contain mostly single-language audio rather than real code-switching
Hi, and thank you for releasing this dataset.
I wanted to report a potential issue with the xtts split.
In several language pairs inside the xtts split, such as Korean-English and Korean-Spanish, the labels/transcripts appear to indicate code-switching. However, when listening to the actual audio, the Korean portions are often completely missing. In many cases, those segments are replaced by silence, or by very abnormal / noisy audio rather than spoken Korean.
This does not seem to be limited to just a small number of examples. In practice, it is difficult to find samples in the xtts split that contain clean and genuine code-switching throughout the utterance.
Because of this, the xtts split does not currently seem reliable for training or evaluation of code-switching ASR / speech models. At least from our inspection, only the read split appears suitable for such use.
It may be helpful to:
- re-validate the audio-text alignment for the xtts split,
- check whether one side of the code-switched utterance was dropped during synthesis or post-processing,
- and possibly provide a cleaned or filtered version of the affected split.
Thank you again for your work. I hope this report is helpful.