Japanese transcription turns speech into Japanese text. If you are comparing speech-to-text tools, start with the scoring method and the kind of audio used in the test. A low error percentage on read sentences does not tell you how a tool will handle your meeting, interview, proper nouns or regional speech.

Japanese transcription is distinct from Japanese-to-English translation and from romanising Japanese into Latin letters. Japan's 2025 romanisation guidance concerns the latter; it does not change speech recognition accuracy. This guide covers Japanese audio to Japanese text and a practical way to evaluate it.

Key takeaways

Why Japanese benchmarks use CER

Written Japanese has no spaces between ordinary words. To compute WER, an evaluator first chooses word boundaries in both the reference and the transcript; two different segmentations can create mismatches even when a reader understands the same sentence. CER compares characters and avoids that segmentation step. The Karita, Sproat and Ishikawa paper reports both measures and explains this limitation.

CER is still an edit-distance score against a reference text. A 5% CER and a 5% WER have different denominators. Neither is a general “95% accurate” guarantee for your audio. When a vendor publishes a Japanese percentage, ask which metric, normalisation, model version and test set produced it. For example, ElevenLabs publishes Japanese WER, while AssemblyAI's multilingual tables mark Japanese as CER; those figures cannot be compared directly.

What alternate spellings change in a score

Japanese permits more than one conventional spelling for some words. A fixed reference can mark a plausible kanji or kana choice as an error. The 2023 study built lattices of plausible respellings of reference transcripts and compared the same recogniser against the original and expanded references. Across three research domains, CER fell by 2.4 to 3.1 percentage points; on its voice-search set, the reduction was 25.16% relative to the original error rate. The recogniser itself did not improve.

Three trained raters assessed 913 proposed spelling variants from the study's YouTube domain. They judged 95.4% plausible, and 16% “great” or better than the original wording. Those are ratings of the proposed variants, not evidence that 95.4% of an engine's errors are harmless. The paper also notes that style and register can affect which spelling fits a particular context. For a transcript you will publish, review names, technical terms and the intended writing style yourself.

How to compare engines on the same task

A vendor benchmark is useful when it discloses its test audio, scoring method, model version and run date. AssemblyAI's current benchmark, for example, separates multilingual FLEURS from Common Voice and explains its normalisation. Its results and model list change, so a copied ranking quickly goes stale; it is also the benchmark operator and a participant. Use published tables to choose candidates, then test them on the same clips from your own workload.

  1. Pick a short, consented sample with the microphone, room noise, speakers and dialect you expect in production.
  2. Write a reference transcript, including the spellings you actually require for names and domain terms.
  3. Run each candidate on the same file and record its model, settings, output format and processing time.
  4. Review meaning-changing words and omissions separately from acceptable kana or kanji variants. Keep the original audio available for disputed passages.

Our transcription-tool comparison covers broader cross-tool testing. This Japanese guide focuses on why the scoring unit and spelling policy matter before you read a ranking.

Dialect and mixed-language recordings

In a 2022 study by Imaizumi and colleagues, a research model trained only on standard Japanese scored 52.5% CER on their dialect test and 14.3% on their standard test. Training that conventional model on both kinds of speech reduced dialect CER to 8.0%, while standard CER moved to 14.9%. The authors also tested dialect-aware methods, including one that improved both sets. These figures describe their datasets and systems; they do not measure MirrorCaption, Kansai-ben specifically or every regional variety.

If a recording contains regional speech, sample that speech in your evaluation instead of relying on a generic “Japanese supported” label. For a recording that alternates between Japanese and English, test the switch points and check which language each segment was written in. A language hint or single selected source language is not proof of reliable code switching. See our multilingual audio transcription guide for the difference between language support and automatic language detection.

What MirrorCaption provides for Japanese

Japanese is a selectable language for MirrorCaption's live transcription and file transcription. Live sessions stream text while speech is in progress; file transcription returns an asynchronous result. Optional translation is a separate step from the Japanese transcript. See live versus file transcription to choose the workflow.

Eligible new accounts can try the free transcription allowance. Current paid options and included hours appear on the pricing page; review them before buying because plan details can change.

Test Japanese transcription on your audio

Try a representative clip, then check names, dialect passages and the export format.

Get Started Free

Frequently asked questions

Why do Japanese speech-to-text benchmarks use CER?

Japanese normally has no spaces between words, so WER requires a word-segmentation choice that can affect the score. CER compares characters without first segmenting words. Both still depend on the reference and normalisation rules.

Does the 25.16% research result mean a quarter of transcription errors are harmless?

No. In one voice-search test set, a 2023 study reduced CER by 25.16% relative to its original score by allowing plausible alternate spellings. That result cannot be applied as a fraction of errors in a different model, dataset or recording.

Can Japanese transcription handle regional dialects?

It depends on the speech and model. A 2022 research study found large differences between standard-only and dialect-inclusive training on its own datasets. Test your actual dialect and review the transcript; MirrorCaption has no dialect selector or published dialect benchmark.

Will MirrorCaption return kanji, hiragana or romaji?

It returns Japanese text without a script selector. Expect normal mixed Japanese writing and review names and specialist terms. MirrorCaption has no furigana or romaji output setting.

Can I export Japanese subtitles as SRT or VTT?

MirrorCaption exports transcript text and JSON, but does not currently export timed SRT or VTT. Use a subtitle editor if you need timed captions for a finished video.

What if my recording switches between Japanese and English?

Test the switch points on a representative file and review each language segment. Selecting Japanese as a file-language hint does not guarantee accurate code switching.