MatterScribe now offers multilingual transcription in 22 languages plus English: it handles recordings that switch languages mid-stream and can add an English translation to any non-English segment. The translation is clearly labeled as machine translation, and the original-language testimony is always preserved as the record. It's live today on every plan.

This isn't just a feature announcement, though. The reason multilingual support looks the way it does in MatterScribe is a failure mode we found in generic AI transcription while building it. It's bad enough that I think every attorney working with non-English recordings should know it exists, and it shaped every design decision we made. To explain it, start with why these recordings are piling up in the first place.

The Language Infrastructure of American Courts Is Shrinking

The right to understand your own trial is constitutional. Courts have grounded interpreter access in the Sixth Amendment's fair-trial guarantee and in due process, and a piece in The Conversation put the problem bluntly: the Constitution promises an interpreter, and US courts often can't deliver one. NPR covered the same crunch in 2025. Demand for court interpreters is up nationwide, judges are delaying hearings, and people are sitting in jail longer while they wait. In Wisconsin, courts billed nearly 30% more interpreter hours in 2023 than five years earlier.

California is the clearest picture of the math. It has the largest court interpreter workforce in the country and roughly 6.8 million residents with limited English proficiency. In one recent fiscal year it logged over 377,000 Spanish-interpreted court events against about 1,310 available Spanish interpreters, and the state just committed $6.8 million over five years to rebuilding its interpreter pipeline.

Court reporters are on the same curve. The US lost about 7% of its qualified court reporters between 2019 and 2023, and federal labor data projects around 1,700 openings a year for the rest of the decade, mostly to replace people leaving the field. So there are more multilingual proceedings every year, fewer people to interpret them, and fewer people to transcribe them. If you practice anywhere with a real immigrant population, you already know what that looks like: hearings postponed because no interpreter was available, transcript queues measured in months, clients waiting in custody for a rescheduled date.

Why "Just Run It Through AI" Makes It Worse

Say you have a recording with Spanish testimony on it and no time or budget for a human transcriptionist. Uploading it to a general-purpose AI tool feels harmless. Worst case, you figure, the Spanish parts come out as garbage and you skip them.

That's not what happens. Modern speech models are trained on so much multilingual data that many of them have quietly learned to translate. Give one a recording where a witness answers in Spanish and it will often just write fluent English at that point in the transcript. No marker. No preserved original. Speaker labels frequently scrambled around the language switch. We call this covert translation.

We didn't read about this in a paper. We hit it in our own testing. Feed English-mode transcription a court-format recording with another language in it, and you get fluent, unlabeled English prose where the non-English testimony was. In one of our test runs it even mistranslated a time reference, turning 9 a.m. into 9 p.m., inside text that gave no hint any translation had happened. Read a transcript like that cold and you'd swear the witness spoke English. Nothing in the document tells you which words were actually said, in which language, or how faithful the hidden translation was.

Think about what that does to a legal record. You can't impeach a witness with a quote they never said in a language they weren't speaking. You can't check a translation against an original that was never written down. And you can't disclose that machine translation was used, because the tool never told you it translated anything.

What Court Guidance Actually Says

The National Center for State Courts has a whole guide on this: Machine Translation: Considerations and Cautions for Courts. It says courts should limit machine translation to brief, low-stakes interactions, treat the output as a starting point for human review, and tell people whenever machine translation was used. Their newer piece on AI in court translation says the same about AI tools specifically, and federal Limited English Proficiency guidance has required meaningful language access in federally funded programs for decades. A tool that translates silently fails every one of those tests at once.

There's an ethics angle for attorneys too. ABA Formal Opinion 500 ties competence and communication to language access. If you can't communicate with your client in a shared language, you need qualified interpretation or translation, and you're responsible for supervising its quality, whether it comes from a person or a piece of software. The ABA's Standards for Language Access in Courts have said much the same to the courts themselves since 2012. And Formal Opinion 512, the ABA's guidance on generative AI, adds that lawyers who use AI tools are responsible for understanding what those tools get wrong and supervising what they produce.

How We Built Around It

When we built multilingual support into MatterScribe, we assumed the transcript would eventually be scrutinized by someone motivated to attack it. That assumption drove five decisions:

The original is the record. Every segment is transcribed in the language actually spoken and kept verbatim. Translation never replaces the original text. It sits next to it, segment by segment.
Translation is labeled, visibly. English translations appear as separate italicized lines explicitly marked as machine translation. You cannot mistake a translated line for testimony.
The disclaimer travels with the document. Every screen and every export carries a notice that translations are machine translations, not certified translations. The disclosure survives forwarding, printing, and filing prep.
A human can check any line in seconds. The Review Dashboard plays the original audio word-synced against the original-language text, so a bilingual reviewer, or the interpreter who was in the room, can verify a line without scrubbing through the recording.
Detection fails loudly, not silently. Recordings run with English as the default, backed by a cross-check that flags non-English speech instead of transcribing it wrong. The exact inverse of covert translation.

What We Measured Before Shipping

It's easy to claim multilingual support and hard to prove it. Here's what our benchmark program actually measured before we shipped:

22 supported languages, plus open-set auto-detect and mixed-language recordings. The pipeline re-transcribes at language switches within a single recording. In validation, one 73-minute interpreter-dense recording split into 49 separate language runs, each transcribed in the language actually spoken.
English protection: the cross-check caught 131 of 132 non-English test files (99.2% sensitivity) and flagged zero of 38 English files, including degraded US appellate argument audio. English recordings don't get spuriously "translated," and non-English speech doesn't slip through.
Translation adequacy: scored against human reference translations on public parallel corpora (FLEURS, Europarl) using chrF, a standard machine-translation quality metric. Translations averaged chrF 64.1 across the supported languages, hit 75.0 for Spanish at full hearing length, and Chinese came in at a 9.1% character error rate on a 61-minute recording.
Playback is word-synced in every supported language, with character-level alignment for Chinese and Japanese and right-to-left rendering for Arabic.
Export parity: DOCX exports carry translation lines 1:1 with the same labeling and the uncertified-translation notice. We checked every line on a staged court-format run: 105 out of 105 came through.

What This Means for Your Practice

The interpreter shortage isn't resolving any time soon, and the recordings keep coming. Here's the division of labor that actually works.

For working access to multilingual audio, use transcription that preserves the original language and labels its translations. That covers reviewing what a Spanish-speaking witness actually said, prepping cross from an interpreted hearing, or triaging hours of bilingual intake recordings. You get same-day access without poisoning the record you'll rely on later.

For the official record, nothing changes. Certified interpreters in the courtroom, certified human translation for filed documents. Machine translation output is not a certified translation. A tool that says so on every page protects you. A tool that translates silently does the opposite.

For client communication, ABA Formal Opinion 500 makes language access a duty, not a courtesy. A labeled, reviewable translation layer gives you a defensible way to work with recordings in your client's language while you line up qualified human interpretation where the stakes call for it.

Multilingual transcription and labeled English translation are live now on every MatterScribe plan. Read how the multilingual workflow handles interpreted proceedings, or start a 14-day free trial and run one of your own recordings through it.

Related reading:

Start your free trial

Upload your first audio file and see a draft transcript in minutes — credit card required.

Start Free Trial

MatterScribe Goes Multilingual: Transcription and English Translation in 22 Languages