Mandarin and Chinese Legal Transcription: Depositions and Recorded Evidence
Mandarin testimony transcribed into Chinese characters, with clearly labeled English machine translation beside it. Recordings that switch between English and Mandarin are handled in one pass.
Transcribing the Recording, Not Interpreting the Proceeding
If you need a Mandarin interpreter for a live deposition, you need a certified human interpreter, and MatterScribe is not that. MatterScribe works on the recording afterwards: the deposition already taken, the hearing already held, the call already sitting in your case file. It turns that audio into a searchable Chinese transcript with clearly labeled English machine translation beside it.
That distinction earns its keep when an interpreted answer is later disputed. Reviewing interpretation problems in international litigation, the Association of Corporate Counsel describes depositions derailed by exactly this: in one case multiple interpreters repeatedly disagreed about the interpretations and left a confusing record, and in another an expert concluded the interpreter had omitted words and phrases and used grammar that made the plaintiff appear evasive. Credibility, not just vocabulary, was on the line.
In each of those situations the recording is the only neutral source. A transcript that keeps the original Chinese next to a clearly marked English rendering lets you compare what the witness said against how it reached the record, line by line and with timestamps. Timing matters as well: under Rule 30(e) a deponent has 30 days from the transcript being made available to submit changes with reasons, and courts scrutinise substantive changes closely. Finding a problem in week one is a different conversation from finding it in month three.
Recordings That Switch Between English and Mandarin
An interpreted Mandarin deposition is a bilingual recording by construction. Question in English, answer in Mandarin, rendering back into English, then an objection in English over the top. Cross-border matters add more: internal calls that move between languages depending on who joined, recorded meetings where a term of art stays in English inside a Mandarin sentence.
MatterScribe divides the recording at detected language boundaries and transcribes each stretch with the model and aligner for that language, then merges the result back in time order. Chinese stretches come out as characters and English stretches as English, within one transcript and one upload. In validation, a 73-minute interpreter-dense recording resolved into 49 separate language runs.
Because each turn keeps its own language tag, speaker label, and timestamp, an interpreted transcript reads as the three-part exchange it actually was rather than a flattened English narrative. That is the structure you need if you ever have to show what the witness said as distinct from what was rendered.
Why Generic Transcription Tools Fail on Chinese Legal Audio
Covert translation. General-purpose speech tools have a documented habit of silently rendering Mandarin speech as English text, with no marker that translation happened and no preserved original. On Chinese audio the cost is higher than elsewhere, because the characters are the only artefact a bilingual reviewer or a certified translator can check the English against. Once the tool discards them, there is nothing left to audit. The National Center for State Courts machine translation guide tells courts to disclose machine translation and treat it as a starting point for human review, and ABA Formal Opinion 512 requires lawyers using AI tools to understand their limits and supervise the output. Neither is possible against a transcript that hides its own translation step.
Homophones and polysemy. Mandarin carries a dense set of homophones, and the mapping from sound to character is genuinely ambiguous in ways English spelling rarely is. A model that picks the wrong character produces text that is fluent, plausible, and wrong. In legal audio the exposure concentrates exactly where it hurts: proper names, company names, place names, dates, and figures.
No word boundaries. Written Chinese does not put spaces between words, so a tool built around English assumptions about tokens, timings, and highlighting behaves oddly on Chinese output even when the characters themselves are right.
How MatterScribe Handles Chinese Recordings
Chinese stays in Chinese. Declare the languages in the recording, or let automatic detection find them, and Mandarin testimony is written as Chinese characters with speaker labels and timestamps intact. The characters are the record and they are never overwritten by their translation.
Character-level playback sync. Because Chinese has no word boundaries, word-level highlighting is meaningless, so playback highlights Chinese transcripts character by character instead. That is what makes a disputed line checkable in seconds: land on the character, hear the audio behind it, and judge the rendering yourself.
Labeled English machine translation. Each Chinese segment gets an English rendering shown as a separate, clearly labeled machine-translation line beneath the original, on screen and in every export, under a visible notice that these are machine translations. The layout is deliberately two-line so that nobody reading the transcript can mistake a translation for testimony.
Full-text search across both. Search runs over the Chinese and the English together, so you can find a passage by the English term you remember and land on the Chinese characters that were actually spoken.
Wrong-language output is flagged, not delivered. A language falling outside the supported set is flagged rather than quietly transcribed with the wrong model, and English-only recordings run a cross-check that surfaces non-English speech instead of mistranscribing it. A confidently wrong transcript costs more than an honest flag.
Mandarin, Cantonese, and Other Chinese Varieties
Mandarin is what we support. Chinese transcription in MatterScribe is built, tested, and measured on Mandarin, and that is the variety our accuracy numbers describe.
Cantonese is not supported. Cantonese, Hokkien, Shanghainese, and Foochow fall outside what the system is built for and may transcribe with materially lower accuracy. This is not a technicality the courts gloss over either: the Consortium for Language Access in the Courts certifies Mandarin and Cantonese interpreters separately, and California certifies them as two of its twelve distinct spoken languages. Both appear independently in the state's interpreter demand rankings, Mandarin second only to Spanish and Cantonese a few places behind, per California court language-access data.
We would rather say this plainly than let you find out on a case file, particularly because Chinese-language evidence does not always arrive labeled by variety. If your recording is Cantonese, run it against trial minutes and judge the output before you rely on it.
Mandarin Depositions, Recorded Evidence, and Business Records
Interpreted deposition recordings. Where a check interpreter was used, or where the parties disagreed on the record about a rendering, an independent Chinese transcript with labeled English gives you something to reason from other than competing recollections. It is also the fastest way to work out whether a disagreement is worth raising at all.
Cross-border discovery and recorded business calls. Litigation and arbitration involving Chinese counterparties produce hours of recorded calls and meetings, most of which do not matter and some of which decide the case. Searchable Chinese plus labeled English lets you triage the volume without listening end to end, and without paying vendor rates to translate material you will never cite.
Recorded statements and intercepted communications. Chinese-language recordings often arrive with a transcript prepared by the other side. Producing your own gives you an independent reading, and shows quickly which passages carry enough weight to justify certified human translation.
Patent and commercial matters with technical vocabulary. Technical terms frequently stay in English inside otherwise Mandarin speech. Because language changes are detected inside the file rather than assumed, those English fragments are transcribed as English instead of being forced through a Chinese model.
What We Measured on Chinese Audio
We report Chinese accuracy as a character error rate of 9.1% on a 61-minute recording. The choice of metric is deliberate. Word error rate, the usual figure for English, requires splitting text into words first, and Chinese has no spaces to split on. Any word error rate for Chinese therefore depends on a segmentation decision that different tools make differently, which makes the number hard to compare and easy to flatter. Character error rate compares the characters directly and means the same thing every time.
Alongside that, translation adequacy averaged chrF 64.1 across the supported languages (chrF is a standard machine-translation quality metric), the non-English cross-check caught 131 of 132 non-English test files at 99.2% sensitivity while flagging zero of 38 English files, and DOCX export parity was verified line for line, 105 of 105 translation lines, on a staged court-format run.
Where the Limits Are
Brief language switches are absorbed. A stretch of one language shorter than roughly six seconds is folded into the surrounding language instead of being transcribed separately. That is a deliberate trade: running eight seconds of Mandarin through an English model and aligner is worse than absorbing a short switch. Sustained turns, which is what interpreted testimony produces, separate reliably. A single English technical term dropped into a Mandarin sentence generally will not.
Machine translation is not certified translation. MatterScribe produces working transcripts and labeled machine translations for attorney review and case preparation. They are not certified transcripts or certified translations. For filings that require certification, engage a qualified human translator. The preserved Chinese transcript gives them an exact, timestamped source to work from, which is more than a silent English-only AI transcript could offer.
MatterScribe does not provide interpretation. We do not supply interpreters and we do not perform interpretation. MatterScribe is software that works on a recording after the proceeding has happened. If you need an interpreter for a live deposition or hearing, engage a certified court interpreter.
Supported Formats and Security
Upload any common format: standard audio (MP3, WAV, M4A, FLAC and more), video (MP4, MOV, AVI, MKV), virtual-meeting exports from Zoom, Teams, and Google Meet, and native court recording formats including .TRM (ForTheRecord™) with no conversion required. All files are encrypted with AES-256 in transit and at rest, processed in US-based SOC 2-compliant data centers, and never used to train AI models.
Get Started
Upload a Mandarin or bilingual recording and see the Chinese transcript with labeled English translation in minutes. MatterScribe's 14-day free trial includes 120 minutes of transcription.
Start Your Free Trial Contact Us
Frequently Asked Questions
Can MatterScribe transcribe a Mandarin deposition recording?
Yes. Mandarin testimony is transcribed into Chinese characters, segment by segment, with speaker labels and timestamps preserved. You can add a clearly labeled English machine translation beneath the Chinese. The original characters are never overwritten, which matters because they are what a bilingual reviewer or certified translator checks the English against.
Can it handle a recording that switches between English and Mandarin?
Yes, in a single pass. An interpreted deposition alternates on almost every turn, and MatterScribe transcribes each stretch in the language actually spoken rather than forcing one language across the file. Chinese stretches come out as characters, English stretches as English, and every segment keeps its own speaker label and timestamp.
Does MatterScribe support Cantonese?
No. Chinese transcription in MatterScribe is built and tested for Mandarin. Cantonese, Hokkien, Shanghainese, and Foochow are not supported and may transcribe with materially lower accuracy. The courts treat this as a real distinction too: the Consortium for Language Access in the Courts certifies Mandarin and Cantonese interpreters separately. If your recording is Cantonese, test it on trial minutes before relying on the output.
Why do you report a character error rate instead of a word error rate?
Because written Chinese has no spaces between words. Measuring word error rate would first require segmenting the text into words, and different segmentation choices produce different scores for identical output. Character error rate compares the characters directly, so the number means the same thing every time it is measured.
Can I get an English translation of Chinese testimony?
Yes. Enable translation and each Chinese segment receives an English rendering shown as a separate, clearly labeled machine-translation line beneath the original characters, on screen and in every export. A visible notice states that these are machine translations.
Is the English translation a certified translation?
No. Translations are machine translations, clearly labeled as such, produced for attorney review and case preparation. For filings that require certification, engage a qualified human translator. The preserved Chinese transcript gives that translator an exact, timestamped source to work from.
Does MatterScribe provide Mandarin interpreters?
No. MatterScribe does not supply interpreters and does not perform interpretation. It is software that transcribes a recording after the proceeding has taken place. If you need an interpreter for a live deposition or hearing, engage a certified court interpreter.