japanese-anime-speech-v2 / dataset_filtering_info.md
joujiboi's picture
Upload dataset_filtering_info.md
e06a93a verified
|
raw
history blame
1.76 kB

Dataset filtering info

Character dedeplication

If any of the following repeat 5 or more times, they were replaced with 4:

'あ' 'い' 'う' 'え' 'お','ぁ' 'ぃ' 'ぅ' 'ぇ' 'ぉ' 'ん'

For example:「ふにゃあああああああ」->「ふにゃああああ」

Reasoning: Some audio contains elongated vowel sounds that are long enough to justify 4 characters

If any of the following repeat 4 or more times, they were replaced with 3:

'~' 'ー'

For example:「きゃわーーーー!?」->「きゃわーーー!?」

Reasoning: Some audio contains elongated vowel sounds that are long enough to justify 4 characters but 3 looks better for 長音符 in my opinion

If any of the following repeat 3 or more times, they were replaced with 2:

'!' '?' 'っ' '!?'

For example:「は!?!?!?」->「!?!?」

Reasoning: These are characters that shouldn't repeat much.

If any of the following repeat 2 or more times, they were replaced with 1:

'…'

For example:「それは……はい」->「それは…はい」

Other filtering

I got rid of trailing「…」

For example:「それは…」->「それは」 and 「…そうです」->「そうです」

I removed「…」that appears before punctuation

For example:「…?」->「?」

Reflection

  • I should consider katakana characters, for some reason I forgot.
  • I think I'll limit「!?」to no repetitions next time
  • There are 192 transcriptions that start with「。」due to string replacements. I should remove these.
  • I should further remove low quality transcriptions who's only characters are: 'ー' 'っ' 'ッ' '…' '!' '?' etc.
  • Open to feedback