How to Fix Japanese TTS Misreadings: The Pronunciation Dictionary Guide
Japanese TTS apps mispronounce fantasy character names and author-invented kanji constantly. Here's exactly why it happens and how to fix it permanently with a per-novel pronunciation dictionary — includes a starter list of common misreadings.
You’ve set up Japanese TTS, found a great web novel, hit play — and the app immediately mangles the protagonist’s name.
“Eirō Arusu” comes out as “Kagerō Arusu.” The hero’s title “Hoshi-yomi” becomes “Sei-ei.” The character’s two-name epithet is read as a completely different compound word. You pause, rewind, wince, and keep going — except now the character sounds wrong in your head for the rest of the novel.
This is the most common frustration with Japanese TTS, and it has a specific cause. Understanding that cause makes the fix obvious.
Why Japanese TTS Gets Proper Nouns Wrong
Japanese TTS engines are statistical models trained on real-world text: news articles, Wikipedia, books, web content. They learn which reading is most probable for a given kanji sequence based on how that sequence appears in the real world.
The problem: light novels and web fiction are full of kanji combinations that have no real-world precedent. An author invents a hero name, a city name, a skill system — and those strings simply do not exist in the training corpus. When the engine encounters them, it falls back to the most statistically likely reading of each component kanji. Which is usually wrong.
There are four common failure modes:
1. Author-invented compound nouns “竜穿大剣（りゅうせんたいけん）” — a made-up sword technique. The engine reads each character separately and produces something plausible but incorrect.
2. Ruby annotations that aren’t recognized
Many web novel sites use formats like 黒猫《くろねこ》 or 影廊（えいろう） to specify readings inline. A good TTS app recognizes these and uses them. But when the source text uses a non-standard format or the ruby gets stripped during import, the annotation is lost.
3. Character names with unusual readings Japanese character names frequently use readings that are uncommon outside of names — or are entirely author-invented. “桐嶋柊也 (きりしましゅうや)” — the 柊 character read as “shuu” is a valid name reading, but the engine has low confidence in it.
4. Series-specific terminology Every long-running series develops its own vocabulary. “概念 (Concept)” used as a game system term, “魔素 (Maso)” as the magic particle name — each series has its own interpretation, and no TTS engine can know them in advance.
The Fix: A Per-Novel Pronunciation Dictionary
The simplest solution is also the permanent one: register the correct reading once, and the app uses it every time that text appears in that novel.
Yomite includes a per-novel pronunciation dictionary. Entries are scoped to one book — so a reading you set for novel A doesn’t affect novel B. This matters when the same kanji means different things in different series.
How to set it up
Step 1: Open the pronunciation dictionary for a book
From the library: long-press the book cover → “Pronunciation Dictionary”
Or from the playback screen: tap the menu icon (•••) in the top right → “Pronunciation Dictionary”
Step 2: Add an entry
Tap ”+ Add”. Two fields appear:
| Field | What to enter | Example |
|---|---|---|
| Text | The exact string as it appears in the novel | 影廊 |
| Reading | The correct reading in hiragana | えいろう |
Tap Save.
Step 3: Verify
Hit play. Use the rewind button to return to the misread passage. Confirm the reading is now correct.
What to Register Before You Start a Long Novel
The most efficient workflow for any novel over 50 chapters:
- Find the cast list or character introduction page (most なろう and カクヨム series include one in the prologue or afterword)
- Register all main character names and their readings before you play chapter 1
- Add world-building terms (magic system names, location names, titles) as you encounter them in the first 5 chapters
This front-loads 5–10 minutes of setup in exchange for dozens of hours of correct playback.
Common Misreading Patterns: A Starter Reference
These patterns appear frequently in isekai and fantasy web fiction. Use them as a guide when building your own dictionary.
Character name kanji with uncertain readings
| Text | Common wrong reading | Likely intended | Notes |
|---|---|---|---|
| 朱雀 | しゅじゃく (shujaku) | すざく (suzaku) | Mythological bird, author often picks suzaku |
| 刹那 | せつな is usually correct | せつな | Verify — this one the engine often gets right |
| 蒼依 | そうい vs あおい | Author-defined | Always check the furigana on first use |
| 夜霧 | よきり vs やきり | Author-defined | Common fantasy name; verify on first appearance |
| 星屑 | ほしくず vs せいさい | Author-defined | Register early if it appears as a title/name |
Skill and world system terms
| Text | Why it gets misread | What to do |
|---|---|---|
| Author-invented skill names | No training data | Register reading on first encounter |
| Unique place names | Compound not in corpus | Register the full place name as one entry |
| Titles / epithets | Often kanji + reading mismatch | Register when the furigana first appears |
When ruby annotations are present
If the source text uses 黒猫《くろねこ》 or similar ruby notation, check whether the TTS is honoring it. In Yomite, automatic ruby capture recognizes this format. If it’s not being applied (you can tell because the reading is wrong even with ruby visible), add the entry to the pronunciation dictionary manually as a backup.
Scope and Limitations
Per-novel scope — Every book has its own independent dictionary. Good for series with unique terminology; slightly more work if you read the same series across multiple imported volumes. For now, the fastest workaround is keeping a text note with your registered pairs so you can re-enter them for volume 2.
Partial string matching — Entries match the exact text string. If a character’s name appears both as 影廊 and as 影廊アルス, you can register both separately, or just register the shorter 影廊 — it will match wherever that string appears.
Hiragana-only readings — Readings must be entered in hiragana. Katakana words (loanwords, sound effects) that the engine misreads should be entered as their katakana pronunciation written in hiragana: e.g., “コンセプト” → こんせぷと.
The Broader Point
TTS mispronunciation is one of the top three reasons people abandon audio novel listening. The other two are voice quality and missing the book in their language. The pronunciation dictionary solves the first. The other two are the reason voice selection matters as much as it does.
Getting the character’s name right from the first chapter isn’t a minor detail — it’s the difference between the voice feeling like a narrator and feeling like an autocomplete error. Five minutes of setup at the start of a long novel is the cheapest investment you can make for the next fifty hours of listening.
Related reading: Aozora Bunko Listening Guide · Web Reader Tutorial
Questions or feedback: [email protected]
— Yomite Team