Voice AI technology

Making Japanese Text-to-Speech Sound Natural on the Phone

How to make Japanese text-to-speech sound natural on phone lines: readings, pauses, speed, honorifics, names and a test cycle that catches mistakes.

Call editorial team · Published · 6 min read

Text-to-speech (TTS) has improved greatly, yet Japanese still trips it up in ways that English rarely does. A caller who hears 日本橋 read with the wrong reading, or an apology delivered in a flat monotone, will not think "the software made a mistake". They will think the company does not care. On a phone line, where audio is already narrow and unforgiving, small errors stand out.

A person wearing a headset listening to a synthesised voice on a phone call
Illustrative image

This guide, written for businesses that operate in Japan or serve Japanese-speaking callers, explains why Japanese is difficult, the six most common problems and how to fix them, and a repeatable improvement cycle. It ends with the tools in Call. If you are choosing between recordings and synthesis, start with Recorded Prompts or Synthesised Voice?.

Why Japanese is hard for TTS

In English, spelling gives a decent guide to sound. In Japanese, the characters often do not. A kanji can have several readings (on and kun), and the right one depends on the word and context. Numbers and counters change their form: the same digits read differently with different counters. Words that differ only in pitch, such as 橋 and 箸, are written differently but sound alike apart from the melody. A system that sees only text must guess.

Six problems and their fixes

What goes wrong and what to do

  • Wrong readingsSpell the reading out in kana for names, places and terms.
  • Missing pausesShorten sentences and place punctuation where a speaker would breathe.
  • Wrong speedSlow down for numbers and instructions; do not speed through the whole script.
  • Wrong voiceChoose a voice whose tone suits the situation, not just the best demo.
  • Wrong politenessUse one consistent level of keigo throughout.
  • Odd proper nounsTest company, product and staff names on their own.

Readings

Most TTS engines support a way to specify how a word should be read, or a dictionary of overrides. Build a list of your own names, place names, product names and jargon, with the correct reading in hiragana or katakana. Check numbers, dates, times and phone-number groupings, since they often read unnaturally. Avoid ambiguous constructions in the script where you can; a rephrase is often easier than a fix.

Pauses and rhythm

Write for the ear. A sentence that is fine on paper may be too long to say in one breath. Break it up, put a comma where a pause belongs, and remember that a phone call needs slightly more space than a page. Lists should have short items. A short pause after the greeting, before the main question, helps callers prepare.

Speed

Slower is not always better. A uniform slow pace sounds patronising; a uniform fast pace is tiring. Vary it: steady for ordinary sentences, a little slower for numbers, addresses and anything the caller must note down. Many callers write while listening, so give them time.

Voice choice

Choose the voice by listening to your own script on a real phone, not only to the demo sentence. Consider the situation: a calm, warm voice suits an apology, a brisk, clear one suits a menu. Male or female is a matter of brand and taste, but consistency across the line matters more. Avoid voices with an unusually bright or emotional tone for serious information.

Politeness

Japanese business phone calls follow conventions of respect. Use one level and keep to it. Mixing casual and formal endings within a message sounds careless, and an over-formal tone can feel cold. Natural openers such as the standard greeting and thanks for calling should be said as a person would say them, not translated from another language. See Greeting Scripts for AI Phone Answering.

Proper nouns

Company names, product names and staff names deserve their own test, since they appear in every call. Say them in a sentence as well as alone. Brand names that mix scripts or contain Latin letters often need a specific reading.

An improvement cycle

From script to a voice that sounds right

  1. Write for the earShort sentences, spoken style, consistent politeness.
  2. Fix readingsAdd names, terms and numbers to a dictionary.
  3. Listen on a phoneCall the line and listen as a caller would.
  4. RateNote every pause, accent or reading that sounds wrong.
  5. ReviseEdit the script or the dictionary.
  6. RetestReplay all earlier fixes to be sure nothing broke.

Keep a regression list: every problem you have fixed, with the sentence that exposed it. Replay it after any change to the script, voice or engine, since an engine update can undo a fix.

Common call moments to check

Some parts of a call are heard by every caller, so they repay extra attention. Listen to each of these on a phone before you listen to anything else.

MomentWhat to checkTypical fix
GreetingCompany name read correctly, warm and unhurriedAdd the name to the dictionary; shorten the sentence
Business hoursTimes and weekdays read clearlyWrite times the way you would say them aloud
Addresses and directionsPlace names and numbersKana readings; slower pace; short chunks
Apology or delayTone is calm, not cheerfulPick a calmer voice; rewrite the wording
ClosingThanks and goodbye sound naturalUse the standard closing phrase, not a translation

A hypothetical shop with a branch named after a local landmark might find that its own address is the most mangled sentence in the whole script. That is common. The words that matter most to the business, its name, its address and its products, are also the ones the engine has seen least. Test them first, fix them first and keep them in the dictionary where the next person will find them.

Before you publish a Japanese script

  • Every name, place and term has a checked reading
  • Numbers, dates and times sound natural in the real sentences
  • Sentences are short enough to say in one breath
  • One politeness level is used throughout
  • The voice suits the tone of the message
  • A native speaker has listened on a phone

Why the phone line matters

Telephone audio is compressed and narrow, which hides some details and exaggerates others. A voice that sounds rich in headphones can be muddy on a mobile line, and certain consonants blur. Always evaluate on a real call, ideally over a mobile network, and again over a landline. Telephone Audio Quality explains why.

Doing it in Call

In Call, the voice your AI agent uses is chosen under Voice, AI voices. Set the language to the one your callers speak, listen to samples on a few cards, press Use this voice, then press Hear what callers hear to listen with the saved settings. The Voice settings and playground area lets you type your own sentence and listen. A phone line can have its own voice, and a line's main language for calls defaults to Japanese; a caller who speaks another language is answered in that language.

The Pronunciation dictionary on the same page teaches the voice how to read names, places and words it gets wrong. Press Add a word, type the word in Written as and its reading in kana in Read as, then type a sentence in Try a sentence to see how it will be read and play it. Only what is spoken changes; the written messages stay as they are. The dictionary holds up to 100 words, and only workspace managers can change it. Finish by testing on a real phone with the agent's test call. See also Language Selection in IVR.

Frequently asked questions

Why does Japanese TTS read some words wrongly?

Kanji can have several readings, and the correct one depends on context. If the system guesses wrong, you can fix it by giving the reading in kana.

Can I control pitch accent?

Some engines allow adjustments, but the controls vary. Often the practical approach is to choose a good voice and rephrase sentences that sound odd.

How do I handle numbers and dates?

Test them in the real sentences you use. Where the reading sounds wrong, spell it out in kana or rephrase.

Which politeness level should I use?

Choose a level that fits your brand and callers, and use it consistently in every message. Mixing levels sounds careless.

Should I use a recorded voice instead?

Recordings sound natural for fixed messages, but changing them is slower. Synthesis suits text that changes often. Many businesses use both.

How often should I check the voice?

After every change to the script, dictionary, voice or engine, and periodically with real calls. Keep a list of past fixes to replay.

Fix Japanese readings with Call's pronunciation dictionary

In the demo console you can try incoming calls in a simulation.

Related guides