Japanese Text to Speech: Pitch Accent and Kanji Readings
Japanese text to speech guide: how pitch accent and kanji readings shape natural speech, plus voice coverage at Azure, Google and Polly in October 2026.
Japanese text to speech depends on pitch accent and kanji readings. Pitch accent tells some words apart, and one kanji can have several readings. As of October 2026, Azure lists 12 Japanese voices and Amazon Polly lists four. Check the words that matter in a test clip before you publish.
What is pitch accent, and why does it matter?
In Tokyo-type pitch accent, the contrast between words lies in the presence and location of a steep pitch drop, called the accent nucleus. A noun with n morae has n+1 possible patterns: unaccented (heiban), initial (atamadaka), medial (nakadaka) and final (odaka). Wikipedia's article on Japanese pitch accent uses the word hashi to show why this matters.
| Kanji | Reading | Meaning | Pattern | |---|---|---|---| | 端 | hashi | edge | Unaccented (heiban) | | 橋 | hashi | bridge | Final-accented (odaka) | | 箸 | hashi | chopsticks | Initial-accented (atamadaka) |
Three words with the same sound are separated only by pitch. A voice that puts the accent in the wrong place can make a sentence sound foreign, or change the word a listener hears. For that reason, listen to the words that matter in your script, not just the overall sound.
Dialects differ too. Many Kansai and Shikoku dialects use contrastive tonal registers across a phrase, such as high-falling, high-level and low-rising, rather than only the accent nucleus. If you need Kansai speech, check whether the voice matches the system you need. The sources reviewed do not show a Kansai-accented Japanese voice.
Research treats this as an open problem. A June 2025 preprint (arXiv 2506.07646) argues that Japanese TTS datasets need phonemic and prosodic labels, and it describes a Whisper-based method for annotating them.
Why do kanji readings go wrong?
Kanji usually have at least two families of readings: on'yomi, the Sino-Japanese reading, and kun'yomi, the native reading. Which one applies depends on the word and the context. According to Wikipedia's article on kanji, the character 生 has eight basic readings, and 今日 is usually read kyō but can also be read konnichi. Furigana, small kana placed beside a kanji, show the reading for ambiguous, rare, literary or non-standard words.
The scale of the problem is large. The Agency for Cultural Affairs (Bunka-chō) says its Joyo Kanji Table, issued as Cabinet Notice No. 2 of 2010 on 30 November 2010, covers 2,136 characters, according to its Japanese language page. Names and rare readings are the cases to test first, because a voice may read a common word correctly and still get a name wrong.
How can you fix a misread kanji?
Use the vendor's pronunciation method where one exists. If none does, write the word in kana. The table lists the methods in each vendor's documentation.
| Vendor | Method | Notes | |---|---|---| | Amazon Polly | Phoneme tag with x-amazon-pron-kana, which encodes pitch accent, or x-amazon-yomigana, which takes conventional kana or Hepburn romanisation | The documentation's example reads 浩一 as ヒロカ'ズ with pron-kana, or as ひろかず or Hirokazu with yomigana. Phoneme tags are fully available for Neural voices and partly for Generative voices | | Azure AI Speech | Phoneme tag using the sapi or ipa alphabets | The sapi phonetic set is defined for ja-JP, and ipa, sapi, ups and x-sampa are the alphabet options | | Google Cloud Text-to-Speech | Custom pronunciation with IPA or X-SAMPA | Available for Chirp 3: HD in ja-JP | | Any tool | Write the word in hiragana or katakana in the script | Editorial advice. Kana is how Japanese writes readings, but it is not a verified feature of every tool |
None of the four Polly Japanese voices is marked Generative, and the phoneme tag is fully supported on Neural voices. Check which Polly voice you are using before you rely on the tag.
Which tools offer Japanese voices?
The table shows what each vendor lists for Japanese, as of October 2026.
| Vendor | Japanese coverage | Notes | |---|---|---| | Azure AI Speech | One locale, ja-JP, with 12 voices | Includes Nanami and Masaru as Dragon HD voices, 7 Standard voices, a Multilingual voice, and two MAI-Voice-2-Flash voices, Haruto and Sakura, which are in public preview | | Google Cloud Text-to-Speech | ja-JP with 30 Chirp 3: HD voices (14 female, 16 male), plus 11 older voices | Chirp 3: HD supports custom pronunciation with IPA or X-SAMPA for ja-JP | | Amazon Polly | 4 voices: Mizuki, Takumi, Kazuha and Tomoko | Phoneme tags with the Japanese kana alphabets | | ElevenLabs | Japanese on Multilingual v2 and Flash v2.5, and in the v3 and v4 language lists | Per-model coverage comes from the ElevenLabs models page | | Gemini API text-to-speech | Japanese listed by name; input language detected automatically | 30 prebuilt voices |
Counts change as vendors add and retire voices, so check each vendor's current list before you commit.
Script tips for Japanese text to speech
- Write hard words in kana when the reading is not obvious from the kanji alone.
- Use punctuation for pauses. Japanese text has few spaces, so commas and full stops do most of the work.
- Test names first. Put the names your audience will hear in a short sample.
- Check English words. The sources reviewed do not verify how a specific voice reads English inside a Japanese sentence.
- Listen to numbers and counters in your sample, and rewrite any that come out wrong.
Japanese text to speech with OndaLabs
OndaLabs has a dedicated Japanese text to speech page. Its 30 prebuilt voices are usable in Japanese, delivery is directed in plain words with supported expressive tags, and the dialogue tool handles two-speaker scenes. Custom voices are available from the Creator plan at $22 per month, which includes 3 custom voices, and can be cloned from a recording you have permission to use or designed from a text description.
The Free plan gives 2,000 credits per month, about 2 minutes on Onda Lite at 1 credit per character, so you can test a name or a hard sentence before paying. Starter at $9 adds Onda Flash, the most expressive model at 1.5 credits per character, plus commercial use. The verified feature list does not include furigana, ruby text or pitch-accent controls, so the kana method above is the one to test.
A simple reading check works like this. Generate a short clip, transcribe it with speech to text, which is listed at 50 credits per minute, and compare the transcript with the reading you intended. The transcript shows what the voice produced, but the sources do not measure transcription accuracy, so listen as well. Compare plans and prices before you choose.
For related reading, see the French text to speech guide and the Hindi text to speech guide.
Frequently asked questions
Which Japanese text to speech sounds most natural?
The sources reviewed do not rank Japanese voices for naturalness, and no independent accuracy benchmark was included. Compare the candidates on your own script. As of October 2026, Azure lists 12 Japanese voices, Google 30 Chirp 3: HD voices plus older ones, and Polly 4 voices.
Why does Japanese TTS read some kanji wrongly?
Kanji often have several readings, and the right one depends on the word and its context. The character 生 has eight basic readings, and 今日 is usually read kyō but can also be read konnichi. Names and rare readings are the first cases to test. Writing the reading in kana removes the guesswork when it matters.
How do I fix a kanji reading in TTS?
Use the vendor's method where one exists: Polly's x-amazon-pron-kana or x-amazon-yomigana, Azure's phoneme tag with the sapi or ipa alphabets, or custom IPA or X-SAMPA on Google's Chirp 3: HD voices. If no method is available, write the word in hiragana or katakana in the script, then listen again.
Can Japanese text to speech read English words?
The sources do not verify how any specific Japanese voice handles English words inside a Japanese sentence. Test a sample that includes the English words you use. Respell any word that comes out wrong, using katakana where that is how Japanese writes the word.
Can I use Japanese AI voices for videos or games commercially?
OndaLabs includes commercial use from the Starter plan at $9 per month. Other vendors set their own terms for each voice and plan, and the sources reviewed do not cover licence terms for Japanese voices. Check the terms before you publish.