Guides
Japanese text to speech, and the kanji problem
Japanese is the language where text to speech either has a real front-end or is guessing.
Last checked 2026-08-14
In short
Japanese needs more than an acoustic model: the same kanji has several readings, and the wrong one is not an accent but a different word. Naadly routes Japanese through a piper-plus voice with a native Japanese grapheme-to-phoneme front-end, so readings and pitch accent come from a real dictionary rather than a guess. Preview names and place names before publishing.
Why Japanese is harder than it looks
- Readings. 生 has more than a dozen. Choosing between them is a dictionary-and-parser job, not an acoustics job.
- Pitch accent. Japanese distinguishes words by pitch pattern. Get it wrong and you are understood but obviously foreign.
- No spaces. The text has to be segmented before it can be read, and a bad segmentation produces a confident, wrong sentence.
- Numbers and counters. 3本 and 3人 are read with different counter words.
What Naadly does
Japanese is served by a piper-plus voice with a Japanese front-end doing the segmentation and reading lookup before synthesis, rather than by a multilingual model that treats Japanese as an accent of everything else. In our own probe measurements that produced materially lower error than routing Japanese through the general catalogue - which is why the other languages that model claims to support are refused rather than shipped half-working.
Checking a render
- Read the proper nouns first. Personal and place names are where dictionaries fail.
- Listen to every number with a counter.
- Give katakana loanwords a listen - they are usually right and occasionally read letter by letter.
- If a reading is wrong, add the word to the pronunciation dictionary in kana and it stays fixed.
Questions
- Can it read kanji correctly?
- Common vocabulary, yes - the Japanese front-end looks readings up rather than guessing. Rare names still need checking, and the pronunciation dictionary is the fix.
- Is Japanese included on the free plan?
- The free plan runs the Piper engine only. Japanese is on paid plans from $15 a month.
- Can I dub a Japanese video into English?
- Yes: transcribe, translate and re-voice in one job. Japanese is among the 35 translation languages.