Chinese, Japanese and Korean artists are almost never identified, and get little or no artist information #709

The problem

StemDeck speaks ten languages, including Chinese, Japanese and Korean. Songs in those languages are almost never identified, and when they are, the artist information is thin or missing.

I measured this on a benchmark of 45 real YouTube uploads across the app's ten languages. The share of songs that found their artist was:

• English: 4 of 4
• Chinese: 0 of 6
• Korean: 0 of 5
• Japanese: 2 of 6
• German, Polish, Portuguese: partly

Titles like these were read as one long song name with no artist at all:

• "周杰倫 Jay Chou【晴天 Sunny Day】-Official Music Video"
• "YOASOBI「夜に駆ける」 Official Music Video"
• "[MV] IU(아이유) _ Palette(팔레트)"

"邓丽君经典金曲 - Teresa Teng's classic songs" found nothing, even though typing "Teresa Teng" into the artist box found her straight away.

Chinese and Japanese artists deserve the same level of respect as Western bands. Right now they get noticeably less.

Why it is not just better title parsing

The titles are part of it. CJK brackets (【】, 「」), full-width separators, underscores and "MV" markers all break assumptions that hold for "Artist - Song". Fixing those alone still leaves several gaps:

• The artist box read Wikipedia only in the UI language, else English. An artist with a long Japanese or Chinese article and a one-line English stub showed no history at all to most readers.
• Native names were never shown. A reader saw only the romanised or translated name.
• Scripts were not understood as the same thing. Traditional and simplified Chinese, full-width and half-width characters, and names written in two scripts ("Jay Chou" and 周杰倫, "IU" and 아이유) were treated as unrelated strings, so a match in one script failed in another.

Constraints for a fix

• Never romanise or guess a name. Only real aliases recorded in MusicBrainz or Wikidata count as the same artist.
• Keep the "no weak guesses" rule. Taking a name from a title alone needs stronger evidence than usual, for example two names in different scripts that both point to one artist.
• No new paid service. The same free sources as for every other language.
• Any character-conversion data carries its licence. A traditional-to-simplified table taken from another project ships with that project's licence file.

Being worked on in the feat/song-identification branch.