Pronunciation
Are mora length, devoicing, counters, and loanwords clean and unambiguous?
OpenRouter · Japanese · female voices
Which voice would you
want to learn from?
Six Japanese-capable TTS models, reading the exact same line. Listen for correct pronunciation, natural rhythm, and whether the voice feels good enough for lessons and voice messages.
TEST SENTENCE
こんにちは。今日は七月二十七日、午後二時半です。東京で新しいコーヒーを買ってから、図書館へ行きましょう。
Hello. Today is July 27th, and it’s 2:30 p.m. After buying a new coffee in Tokyo, let’s go to the library.
The listening room
Use headphones. Replay tricky words like 今日, 七月, and 図書館.
How to judge
Don’t only pick the prettiest voice.
Naturalness
Does it sound like a Japanese speaker talking, rather than carefully reading symbols?
Teaching fit
Is it clear, warm, and steady enough to imitate without becoming stiff or robotic?
Why only six?
OpenRouter listed 15 speech models when this test was made. These six had a female preset voice and documented or provider-confirmed Japanese support. Qwen Audio 3.0’s included voices are Chinese/English only; Microsoft MAI-Voice-2 currently exposes no Japanese voice on OpenRouter; Voxtral’s presets are English/French; and Zonos, Orpheus, and Sesame are English-only. They were left out instead of producing misleading “Japanese” samples.
Generated 27 July 2026 through OpenRouter’s dedicated speech API. Audio is a single generation per model, not a scientific benchmark. Prices and available models can change.