Turn Japanese meetings, interviews and lectures into searchable text with speaker labels — kanji, hiragana and katakana as they should be written.
iOS and Android. 30 minutes free, no card.
Does SozAI transcribe Japanese?
SozAI transcribes Japanese audio and video into text using kanji, hiragana and katakana as appropriate, with automatic speaker labels and word-level timestamps. It exports to TXT, PDF, SRT and VTT. Recordings travel over HTTPS and are stored on our servers in Germany, and every account gets 30 minutes, once per account, without a card.
日本語, written the way it is actually written.
Specification
| Native name | 日本語 |
|---|---|
| ISO 639-1 code | ja |
| Writing system | Kanji, Hiragana, Katakana |
| Text direction | Left to right |
| Speakers worldwide | ~125 million |
| Varieties covered | Standard (Tokyo), Kansai |
| Speaker labels | Yes — automatic diarization |
| Timestamps | Word level |
| Summaries | Yes, in the same language |
| Export formats | TXT, PDF, SRT, VTT, JSON |
| Maximum file size | 500 MB |
| Free tier | 30 minutes once, no card |
| Where data is stored | Stored on our servers in Germany, encrypted in transit |
| Public transcripts in this language | 483 |
Workflow
Record straight into the app, upload a file up to 500 MB, or paste a YouTube link. SozAI reads m4a, mp3, wav, aac, mp4 and mov.
Choose Japanese, or leave detection on. Recordings that mix Japanese with English technical vocabulary stay in a single transcript.
The transcript arrives with speaker labels and word timestamps. Search it, ask it questions, generate a summary in Japanese, or export SRT and VTT subtitles.
Field notes
Japanese is written without word boundaries, so the transcript has to decide where words begin and end before it can time-align them or make them searchable. Two segmentations of the same sound sequence can both be valid Japanese with different meanings. This is why word-level timestamps in Japanese are a harder claim than in a spaced language.
The same word can be written in kanji, hiragana or katakana, and the choice signals register, emphasis or foreignness. Writing everything in hiragana is technically readable and looks like a child wrote it; over-applying kanji reads as stiff. A usable Japanese transcript has to make the same choices a person would.
Japanese has an unusually dense field of homophones — kikan, koutai, seikou each map to many different kanji spellings. The audio contains no information about which one was meant. Selecting the right characters is a language-modelling problem, and it is where most remaining Japanese transcription errors live.
Honorific, humble and plain registers use different verbs, not just different endings, and Japanese business recordings move between them constantly. A transcript that flattens keigo is not merely less polite — it misrepresents who was deferring to whom, which matters in a meeting record.
From the archive
483 public Japanese transcripts so far — every one produced by the same engine described above. Read them in full, no account needed.
卵の8倍！？この食べ物は卵よりもたんぱく質が多い ― 60歳を過ぎても自然に筋肉を守る方法【食で長生き】 — Transcript
【漫画】「このバカ息子が！」感謝の気持ちを込めた手料理を振舞ったら嫁が泣きながら出ていった… — Transcript
なぜ平沢進のツイッターアカウントはあんなに気持ち悪いのか？ — Transcript
【第60回】MARIKA MC回！チート級メンバーで挑む賞金100万円のカラオケBINGO🎯 by シアーミュ… — Transcript
心が疲れている12のサイン【ストレス】【体が動かない】【心が動かない】 — Transcript
【月額9万円】僕のAIの使い方、全部見せます。 — TranscriptSozAI transcribes 99 languages. These have a page of their own.
Answers
Short answers you can quote. Longer ones live in the sections above.
It uses kanji, hiragana and katakana the way a person would write them, rather than defaulting to all-kana output. This is what makes the transcript readable at normal speed and searchable — kana-only Japanese is technically decodable but nobody reads documents that way.
Yes. Diarization runs on the audio rather than the language, so each speaker gets a label and every turn a timestamp. For 議事録 and interview records this is what turns a recording into a document you can circulate.
Accuracy is highest on clear, single-speaker standard Japanese and drops on fast overlapping conversation, heavy Kansai speech and poor phone audio. The most common residual errors are homophone choices — the sentence structure is right and one word is written with the wrong kanji, so proofreading is worth budgeting for.
SozAI exports SRT and VTT in UTF-8 with word-level timing, so Japanese text loads correctly in YouTube, Premiere Pro, Final Cut Pro and DaVinci Resolve. Line breaking still deserves a human pass, since Japanese subtitle conventions differ from Latin-script ones.
Every account gets 30 minutes of transcription once, at no cost and without a card. Beyond that, Premium is a subscription through the App Store or Google Play — from $3.99 a week or $44.99 a year in the US — and the price is the same in every language. Current plans are listed on the pricing page.
Where to use it
Everything below works in Japanese, with the same speaker labels and word timestamps.
Transcribe your first Japanese recording
Record, upload a file or paste a link. Speakers, timestamps and a Japanese summary come back together.
Free on iOS and Android. No account required for YouTube transcripts.
Point your camera
Scan the code to open the download page on iOS or Android. 30 minutes of transcription free, no card.
