How do you get hiragana readings from anime song lyrics, and can you trust them?
Author: SunClaw Team · Updated: 2026-11-17
The short answer
No. Do not treat them as a standard reading. In UtaNote, the kana, romaji, translation, and word-by-word breakdown of a song are a study draft written by a language model. They are not a publisher's checked lyric sheet. The prompt tells the model to be accurate. The model can still be wrong. When search is not sure of a field, the prompt allows a reasonable guess.
This page does not print the "correct" kana for any song. Doing that would treat one model draft as a fact.
What the lyric draft contains
The parse is one model call. The model is qwen3.5-plus. The lyric call uses temperature 0.2. The prompt asks for JSON with:
- Each line: the Japanese text, kana for the whole line, romaji, a Chinese translation, and grammar notes.
- Each word: the surface form, kana, romaji, a Chinese gloss, part of speech, an optional grammar note, the conjugation if it is a verb, and an optional JLPT level from N5 to N1.
- Cultural notes and a short grammar summary for the song.
The same prompt states two limits you should believe. First, it asks for only the first 4 to 6 lines, as a sample, not the whole song. This screen is not a full lyric sheet. If the model returns fewer lines, more lines, or stops halfway, you see what it returned. If the reply is cut off and does not end as a finished JSON object, the app reports an error and tells you to retry or pick a shorter song. It does not silently fill in the rest of the track.
Second, the prompt says the lyrics must be the real lyrics and the kana and romaji must be accurate. That is an instruction to the model. It is not a check the app runs against an official lyric book afterward. There is no lyric library in the app to check against.
Search guesses too
Search returns up to 5 hits, with title, artist, anime, type, and a match score. Type has to be one of OP, ED, insert song, character song, theme, or other. The prompt says that if a field cannot be determined, a reasonable guess may fill it.
The anime title or the artist on a hit can be a guess. Generating a lyric draft after that is a second model pass on top of a title that may be wrong. Neither step looks up a checked songbook.
The kana chart is a different thing
The kana chart in the app is a written table. It is not the output of the lyric model. Basic gojuon, dakuten, handakuten, and youon are on the chart, with mnemonic hints, and a tap can play the sound. The achievement for the basic set counts 46. The achievement that includes dakuten and youon counts 100.
Learn how あ is pronounced from that chart. For the kana of one lyric line, treat the model draft as a draft: compare it with what you hear, and correct your own notes when it is wrong. Do not copy a line of kana off the screen into homework as if a teacher had checked it.
FAQ
Q: Is the kana the official reading of the lyric? A: No. A model writes it from the prompt. The prompt asks for accuracy. The app does not check the result against a lyric library.
Q: Why are there only a few lines? A: The prompt asks for the first 4 to 6 lines. It is not a full-song lyric sheet.
Q: Is the song search always the right title? A: Not always. A field the model cannot settle may be filled with a guess. Before you generate a lyric draft, check that the title and artist are the song you meant.
Q: Can the kana chart be randomly wrong as well? A: The lyric draft can. The kana chart is a separate written table. Use it to learn the characters. Do not use it as the reading of one lyric line.
Use it as a draft
Confirm the kana you do not know on the chart first. Then read the few lines of the lyric draft. When the sound you hear and the kana disagree, trust the source you have checked. Do not trust that one model pass.