Japanese Live Transcription Evaluation

Japanese Speech-to-Text Accuracy: 4.33% Word Error Rate

Japanese is one of the hardest languages for automatic speech recognition: three writing systems, no spaces between words, and many homophones. We tested Conference Captioning on real Japanese speech and on a second, reproducible real-time test, and both landed at roughly 95% accuracy or better.

4.33%

Word Error Rate

On a real 254-word speech sample. Below 10% is considered good and below 5–6% is top-tier for professional use.

95.25%

Word Information Preserved

The transcript kept over 95% of the original information, with only 4.75% lost.

<1.5s

Real-Time Latency

In the live test, captions appeared with less than 1.5 seconds of lag, suitable for presentations, classrooms and conferences.

Why Japanese Is Hard for ASR

  • Complex writing system: Kanji, Hiragana and Katakana are mixed in one sentence.
  • No word delimiters: the engine must segment sentences into words itself.
  • Homophones: for example, きかい can mean 機械 (machine), 機会 (opportunity) or 奇怪 (strange).
  • Kanji variants: 温かい and 暖かい both mean “warm” with subtle differences in nuance.
  • Fillers and colloquial speech: あのー and えっと must be handled correctly.

Evaluation Methodology

Both tests used the Multilingual speech model, which can switch languages in real time and is not our most accurate model for a single language.

  • Test 1 (October 2025, app v2.15): a clear public-speaking sample transcribed by Conference Captioning and compared with a hand-made reference transcript.
  • Test 2 (app v2.13.0, MacBook Pro M4): a two-minute Japanese speech written for high school students, converted to audio with Google Translate’s text-to-speech, and transcribed live with no pre-uploaded data.
  • Metrics: Word Error Rate (WER) = (substitutions + deletions + insertions) / words in the reference, and Word Information Preserved (WIP).
Live ASR
Japanese Speech
Word Error Rate
Reproducible

Test 1: Real Speech Sample

Error Tally

TypeCountExamples
Substitutions (S)5温かい→暖かい, 5年→お念, 風→ふう, ピタッと→来たとこ, 未来→見らえ
Deletions (D)4が, です, ね, に
Insertions (I)2あのー, でも
Words in reference (N)254 

WER = (5 + 4 + 2) / 254 = 11 / 254 ≈ 4.33%

What the Errors Show

  • Few semantic errors: 温かい→暖かい and 風→ふう are understandable and often acceptable in informal text. The larger misses were rare phonetic slips.
  • Verbatim accuracy: fillers such as あのー were captured, which matters for legal and research transcripts.
  • Natural speech: というのは was written as the more conversational ってのは, showing the model follows spoken Japanese, not just formal writing.

Reference (Ground Truth)

Conference Captioning Output

Test 2: Real-Time Text-to-Speech Test

95.8%

Overall Accuracy

About a 4.2% word error rate over roughly two minutes of speech, transcribed in real time.

Sample Differences

ReferenceOutputType
未来を片づくります未来を形作りますSubstitution
上熱を持って情熱をもってSubstitution
ら心盤羅針盤Substitution
お忘れないこと忘れないことDeletion
ありがとうございましたありがとうございましTruncation

The reference text itself contained the odd wording shown on the left, so several “errors” are the model correcting it. None affected meaning.

Original Speech Text

Conference Captioning Output

You can reproduce this test: the speech was generated with ChatGPT using the prompt “Give me a 2 minute inspirational speech written in Japanese for high school students to motivate them to study hard and be passionate about what they do”, played with the Listen feature in Google Translate, and fed into the app.

Key Findings

  • Both tests came in at roughly 95% accuracy or better (4.33% and about 4.2% word error rate).
  • Errors were mostly small substitutions or missing characters, not changes in meaning.
  • Captions were effectively instant, under 1.5 seconds of lag.
  • This used the Multilingual model, which switches languages in real time, so a dedicated single-language model has room to be even more accurate.

FAQ

How accurate is Conference Captioning for Japanese?

Our tests measured a 4.33% word error rate (95.25% information preserved) on real speech and about 95.8% accuracy in a live real-time test.

Does it work in real time?

Yes. Captions appeared with less than 1.5 seconds of lag, with no pre-uploaded audio.

Can Japanese speech be translated for other attendees?

Yes. A speaker can talk in Japanese and attendees can read live captions or translations in their own language.

Related Evaluations