Ciji Transcribe

詞記逐字稿

Transcripts for Mandarin meetings — the kind where English technical terms land in every other sentence. The audio is processed on your device, and the output is Traditional Chinese: 網路, not 网络; 軟體, not 软件. Bring a recording from whatever you already use.

It remembers what your terms get misheard as Not a spell-checker — a homophone repair list that grows as you use it.

Traditional Chinese, held steady Speech models drift to Simplified. A deterministic table pulls it back.

Nothing leaves the device No account, no cloud transcription service, no per-minute billing.

Any recorder works Phone, watch, dedicated recorder, a file a colleague sent you.

The part nobody else does

An English word-replacement list cannot fix a Mandarin homophone. Measured on a real meeting: with prompting alone the term came back wrong every single time — and always the same way, which is exactly what makes it repairable. The repair pass recovered all fifteen occurrences.

甲乙丁→甲乙丙
甲乙戊→甲乙丙
Widgit→Widget

Deliberate placeholders. The entries of a real glossary are, by construction, the words one particular meeting actually used — which is why yours stays on your device and nobody else's belongs on this page.

A glossary that learns

After each meeting the app mines the transcript for likely mishearings and asks — "Widgit — did you mean Widget?" One tap accepts, one dismisses. Suggestions never apply themselves, and once a term is in, it is fixed in every future transcript.

Traditional Chinese, enforced

Speech models trained mostly on Simplified text drift there under load. A 171-character and 72-phrase table converts the output back — and knows that 公里 is not 公裡 and 批准 is not 批準, which a character-by-character swap gets wrong.

Check it against the audio

Every line carries its start time. Tap a segment to hear exactly that stretch of the recording — because a transcript you cannot check against the audio is a transcript you have to trust.

What you actually get

ImportBring in a recording from the share sheet — phone, watch, recorder, or a file someone sent you.
TranscriptionRuns on the device. The speech model is downloaded when you choose it, and deleting it frees the space back.
GlossaryOne term per line, with the wrong forms it gets heard as. Applied to the transcript, and reused in every meeting after.
SuggestionsCandidate mishearings mined from each transcript, accepted or dismissed in one tap.
ReviewTimestamps on every segment, and per-segment playback to check the text against what was said.
ExportPlain text, and subtitles.
SpeakersTelling apart who spoke is far more expensive than transcribing what was said — so it comes later, not in the first release.

Not released yet — this page describes what is being built.