Chinese Transcription for Mandarin Recordings

Chinese transcription on this page runs speech recognition with the language fixed to Chinese. You get a transcript written in Chinese characters, split into timed segments that play back against the audio and become caption cues. Segments are joined without the stray spaces that appear when Chinese text is stitched together like English words.

Spoken numbers usually come back as digits, dates as 2026年10月8日, and percentages with a percent sign. English product names and acronyms spoken inside a Chinese sentence stay in Latin letters, but they are also a common source of mistakes. Punctuation may be half-width, so tidy commas and question marks before you publish.

The transcript is not translated, summarized, or converted between simplified and traditional characters, and the model does not always pick the same script: in our test, adding background noise switched a Mandarin file to traditional characters. This is Chinese speech to text for people who need the original-language record: a researcher checking a Mandarin interview, an editor preparing captions, or a team member who wants a colleague fluent in Chinese to review a call.

Mandarin is the tested case. Cantonese is a separate choice in the language list; our only check was one synthetic sentence, where the Cantonese setting wrote colloquial Cantonese characters and still misheard one word. We have not tested Hokkien, Taiwanese, Shanghainese and other Wu varieties, Sichuanese, Hakka, or heavy regional accents. Try a short excerpt before relying on those recordings.

Names are the weakest point. Many Chinese given names share a sound with other characters, so the model picks a plausible spelling rather than the person’s real one, as with 林晓雨 in the sample. Add important names and terms in the settings before submitting, keep a list of the correct characters, and confirm company names, places, and English acronyms against your notes.

Silence and pure noise are a known trap for this model: it can write a subtitle credit or a sign-off that nobody said. The tool removes those lines and fails the task with a clear no-speech message when no spoken words remain, rather than saving a full fake transcript. A short invented phrase inside a long, noisy recording can still slip through, so listen to quiet passages.

Upload MP3, WAV, M4A, AAC, FLAC, OGG, OPUS, WebM, WMA, MP4, or MOV, or record in the browser. Each file can be up to 1GB and 60 minutes. Split a longer recording into parts; each part becomes its own task and transcript.

Processing runs in the cloud and uses account credits based on duration. The estimate appears before you submit, and the server checks your balance. Read plans and usage limits for current terms. Only the task owner can open the saved transcript and its exports.