Native Japanese workplace test

Can AI Understand Japanese Workplace Homophones?

We tested eight same-sounding words inside complete workplace sentences. Notta and Sonix tied for first, but every tool confused producing creative content with manufacturing a physical object.

45-second recordingSame audio for all three toolsChecked by a native speakerTested August 25, 2026

See the results ↓Read Test 06A

Independent hands-on test & affiliate disclosure

We wrote and recorded the Japanese source, uploaded the same M4A file to every tool, and checked each sentence manually. Notta and Sonix links are affiliate links: we may earn a commission after a qualifying purchase, at no extra cost to you. This does not affect our scores.

IN THIS TESTEight target words Results Shared failure Why two scores Method & limits FAQ

01 / TARGET WORDS

Four homophone pairs in natural workplace context

The pronunciation alone was not enough. Each service had to use the subject, object, and surrounding action to select the intended kanji.

会う・合う

meet a person · fit a foot

変える・帰る

kaeru

change a time · return home

直す・治す

naosu

correct a typo · cure a cold

制作・製作

seisaku

produce creative content · manufacture a physical object

THE SOURCE

One 45-second recording made by a native Japanese speaker in Japan

The sentences covered meeting a staff member, shoe fit, changing a meeting, returning home, proofreading, recovering from a cold, producing an ad video, and manufacturing a wooden chair.

02 / RESULTS

Notta and Sonix tied for first

We report both the target-kanji score and the full-sentence score. This prevents an unrelated recognition error from disappearing behind a seemingly strong homophone result.

Tool	Target kanji	Full sentences	Meaning errors	Observed error
Notta	7/8	7/8	1	Used 制作 instead of 製作 for manufacturing a wooden chair.
Sonix	7/8	7/8	1	Made the same 制作／製作 substitution as Notta.
TurboScribe	7/8	6/8	2	Also changed 誤字 (typo) to 5時 (5 o’clock), altering the proofreading instruction.

Target	Intended meaning	Notta	Sonix	TurboScribe
会う	Meet a new staff member	Correct	Correct	Correct
合う	Shoes fit the speaker’s feet	Correct	Correct	Correct
変える	Change the meeting start time	Correct	Correct	Correct
帰る	Go straight home after the meeting	Correct	Correct	Correct
直す	Correct a typo before publishing	Correct	Correct	Target correct; 誤字 became 5時
治す	Recover from a cold	Correct	Correct	Correct
制作	Produce an advertising video	Correct	Correct	Correct
製作	Manufacture a wooden chair	制作	制作	制作

Read the product details: Notta review, Sonix review, TurboScribe review.

03 / SHARED FAILURE

Every tool chose the creative-content kanji for a physical chair

INTENDED

新しい木製の椅子を製作するため、職人が図面を確認しています。

The craftsperson is checking a drawing in order to manufacture a new wooden chair.

ALL THREE OUTPUTS

新しい木製の椅子を制作するため、職人が図面を確認しています。

制作 is normally used for creative works such as videos, art, or advertising. 製作 is the intended choice for making a physical chair. The output remained readable, which makes the error easy to miss.

This was not an isolated two-word prompt. The sentence included three strong physical-production clues: 木製の椅子 (wooden chair), 職人 (craftsperson), and 図面 (drawing or plan). The shared miss suggests this pair deserves manual review in professional transcripts.

04 / SCORING

A homophone score alone can hide another serious error

TARGET WORD

TurboScribe correctly wrote 直す for correcting a typo.
That target therefore counted as correct.

FULL SENTENCE

It changed 誤字 (typo) to 5時 (5 o’clock).
The proofreading instruction no longer made sense.

That is why TurboScribe received 7/8 for target kanji but only 6/8 for full-sentence meaning. For business use, checking the whole sentence is more useful than checking a curated keyword alone.

TRY THE TOOLS

Test your own Japanese audio before paying.

Use audio whose intended wording you can verify line by line.

Try Sonix free →Try Notta →

05 / METHOD & LIMITS

How we ran the comparison

1
One native-recorded source.
A single speaker read the final eight-sentence Japanese script.
2
Exactly the same M4A file.
Notta, Sonix, and TurboScribe received the same 45-second recording.
3
Japanese processing selected.
TurboScribe used Whale mode; the other tools used their available Japanese transcription flow.
4
Native manual review.
Every output was checked against the intended Japanese script.
5
Narrow conclusion.
One recording cannot establish a universal accuracy percentage or permanent winner.

06 / FAQ

Japanese workplace homophone FAQ

Which tool was best in this workplace Japanese homophone test?

Notta and Sonix tied for first. Each selected 7 of 8 intended target kanji and preserved 7 of 8 full-sentence meanings. TurboScribe also scored 7/8 on the target kanji, but an additional 誤字-to-5時 recognition error reduced its full-sentence score to 6/8.

Can Japanese transcription AI distinguish 制作 from 製作?

Not reliably in this test. All three tools wrote 制作 for a sentence about a craftsperson manufacturing a wooden chair, even though the sentence also mentioned wood, a craftsperson, and a drawing.

Why score target kanji and full sentences separately?

A tool can select the target homophone correctly while making another error in the same sentence. TurboScribe preserved 直す but changed 誤字 to 5時, so its target-kanji score alone would hide a meaning-changing mistake.

Does this test prove one tool is always more accurate?

No. This was one 45-second recording from one native Japanese speaker. It reveals a specific weakness worth checking, not a universal accuracy rate for every voice, room, or topic.