Independent testing from Japan
We test transcription tools with real Japanese audio, then a native Japanese reviewer checks every number, name, and kanji choice by hand.
Updated September 5, 2026 · Eight controlled recordings
CONTEXT TEST
橋の端で箸を落とした。
Preliminary—not a final winner. More real-world tests are coming.
01 Recorded in Japan
02 Reviewed by a native speaker
03 Same audio for every tool
04 Weaknesses published
Early benchmark results
These are hands-on results from eight short controlled recordings. The new workplace test exposed a shared 制作／製作 error even with strong context.
The Notta and Sonix buttons below are affiliate links. We may earn a commission after a qualifying purchase, at no extra cost to you. Other product buttons are direct links.
MEETING & NOTES
Strongest overall across eight recordings, but our new workplace homophone test exposed a 制作／製作 meaning error.
Watch out: It removed hyphens from A-2048 and XR-205, wrote numbers in kanji, and rendered 青葉 phonetically as アオバ.
AUDIO TRANSCRIPTION
Produced readable numbers and codes, but the noisy two-speaker test exposed serious proper-noun and instruction errors.
Watch out: In noise, 八潮配送センター became 八代海藻センター and 取扱注意 became 無限使い注意.
AUDIO TRANSCRIPTION
Excellent speaker labeling in quiet audio, but controlled noise collapsed every turn into one speaker and changed a quantity.
Watch out: In noise, 14箱 became 4箱 inside a correction and every voice was labeled Speaker 1, despite 97.69% very-confident words.
MEETING & NOTES
Recognized ミライテック and two of three homophones, but changed a price. Later tests were blocked by the Basic plan's import cap.
Watch out: 1万2,800円 became 12,008百円—a materially different amount if read literally.
WEB TRANSCRIPTION
The no-signup web converter accepted all five recordings but failed during processing every time.
Watch out: This is a web-tool reliability finding, not an accuracy verdict. The signed-in app may behave differently.
Biggest lesson so far
Polished text can still be confidently wrong.
Sonix reported 95.83% confidence while changing A-2048 to A 248 and XR-205 to エックス 2005. TurboScribe’s clean-looking output changed 仕様 to 使用, while JotMe failed to return text for five different files. In controlled noise, Sonix also lost speaker separation and TurboScribe changed two important terms. Native review still matters.
Our method
Every tool receives the same original audio. We separate recognition mistakes from punctuation and formatting, and we do not score a tool that fails to produce a transcript.
Read our full scoring standard →A short, naturally paced recording used to check basic content accuracy and punctuation.
Dates, prices, an address, an invented company name, ChatGPT, and 橋・端・箸 in context.
Fillers, four self-corrections, shortened phrases, times, responsibility, and agenda details.
Across quiet and noisy takes, Notta labeled all 18 scripted turns correctly. Noise collapsed Sonix to one speaker; TurboScribe labeled 8/9 noisy turns correctly. TurboScribe's quiet take remains pending.
Honorifics, A-2048, XR-205, quantities, prices, a delivery-date change, API, CSV, and a reply deadline.
Read the full test →Two focused recordings tested everyday actions and workplace language, including 採る, 贈る, 直す, 治す, 制作, and 製作.
A transcript can look polished to a non-Japanese speaker while getting names, kanji, politeness, or meaning wrong. Native Japanese AI Lab is built in Japan to make those mistakes visible—and help international users choose with confidence.
Some product buttons are affiliate links, and each one is labeled beside the button. We may earn a commission after a qualifying purchase, at no extra cost to the reader. Compensation never changes our test files, scoring method, ranking, or the weaknesses we publish.
