English and Music in TEDxJP 10k corpus

[eiichiroi]

The TEDxJP 10k corpus contains some inapporopriate data for evaluation of Japanese speech recognition.

In following videos, Japanese people talk in English.
But the corpus uses subtitles that are automatically translated into Japanese.

• Aj-DXM5Zqms
• Ba5Jl1_JKZY
• Mc044I55SCY

Then, following videos contain some music (corpus uses the interval of music).

• ydhfjNRFzaM
• BLElQZfR_2M
• TWeYkdIQsk0
• Ab-KZT06gR0
• gffgHgnEhtA
• kU9LcoHaFLo

It may be better to be removed from the corpus.

[hfujihara]

@eiichiroi
Thank you for pointing out this issue.

We examined the videos you mentioned and

• removed English utterances in Aj-DXM5Zqms, Ba5Jl1_JKZY and gffgHgnEhtA
• kept Mc044I55SCY as a part used in our data includes only Japanese utteranes
• Kept ydhfjNRFzaM, BLElQZfR_2M, TWeYkdIQsk0, Ab-KZT06gR0, and kU9LcoHaFLo as, though they are singing voices (or rap), their transcriptions are actually correct.

[hfujihara]

The above modification was made in https://github.com/laboroai/TEDxJP-10K as TEDxJP-10K_v1.1 .

[eiichiroi]

Thank you!