Datasets:
Speaker diversity information
Dear Sardor, many thanks for your outstanding work!
Could you clarify the diversity of voices in this and relative datasets (it_youtube_uzbek_speech_dataset, news_youtube_uzbek_speech_dataset), please? For example, in podcasts we may have a podcast host who will appear in each episode, does it mean that his/her voice can appear in 50% of data samples? Or does "intelligently filtered" in the dataset description imply that it was balanced somehow?
Would appreciate your answer
Hello, thanks!
Actually, this podcast is more in an interview format. Yes, there is a host, but he mainly asks questions while the guest does most of the talking. Also, it wasn’t taken from just one channel. I think the host's audio makes up about 10–15% of the youtube_uzbek_speech_dataset. But there are no diversity metrics available. I assume it’s not a big issue, as when mixed with other public datasets, it should be fine.
As for the news_youtube_uzbek_speech_dataset — it has pretty good diversity since it was collected from different sources and includes various journalists.
Brilliant, thank you for sharing this information!