The Court Transcript Dropped “Ex.” The Interpreter Caught It.

The word on the printed transcript was “wife.” Elena had said “ex-wife.”

She circled the word with a pen after the hearing, then wrote the missing prefix in the margin. The sheet became her record of a problem that had taken less than a second to create and several minutes to explain.

Elena is a composite of court interpreters who described similar work. She had interpreted Spanish and English in state courts for 11 years when the courthouse began testing live automated transcription. The text appeared on a monitor a few words behind each speaker. Staff described it as a backup, not the official court record.

At first, that distinction made sense to her. A rough transcript could help staff find a point in a long hearing, and it could give some people another way to follow speech. The system handled dates and common courtroom phrases better than Elena expected, especially when one person spoke at a steady pace into a working microphone.

Then the assignment changed. Interpreters were expected to watch the transcript and flag mistakes while they worked.

That meant Elena had to listen to Spanish, hold a complete thought in memory, render it into English, watch the English words appear on a screen, compare them with what she had just said, and decide whether a difference mattered enough to interrupt the proceeding. No task disappeared to make room for the new one.

Interpreters who contributed to this composite called the added task “second listening.” The phrase is plain, but the work is not. Simultaneous interpreting already divides attention between incoming speech and outgoing speech, while consecutive interpreting depends on memory, notes and the ability to recognize when a speaker has finished a thought. The transcript added another stream that arrived late and sometimes changed as new words came in.

The word on the page

The hearing concerned an alleged violation of a protection order. A witness referred in Spanish to “mi exesposa.” Elena interpreted the phrase as “my ex-wife.” On the monitor, the transcript showed “my wife.”

The missing sound was short. The interpreter’s “ex” fell near a rustle of paper and the start of another person’s speech, while “wife” remained clear. Automated speech recognition does not listen by pausing over legal significance. It converts slices of sound into possible word sequences, then ranks those sequences using patterns learned from speech and text.

A common phrase with strong audio can outrank a less common phrase containing a weak syllable.

The resulting sentence looked finished. There was no blank, warning or visible confidence score beside “wife.” Punctuation had also been inserted, giving the line the appearance of settled prose even though the system was producing a draft from live audio.

That confidence was visual, not legal. “Wife” and “ex-wife” describe different relationships, and the distinction could affect how a listener understood the history between two people, their household and the timing of events. Elena did not decide what the difference meant under the law. Her job was to preserve the speaker’s meaning.

She noticed the error because she happened to look at the monitor after finishing the sentence. She raised it, the line was discussed, and the proceeding slowed while the record was clarified. Later, she printed the relevant sheet and circled “wife.”

The correction was treated as proof that the backup worked. Elena saw the same event differently. The system had not caught her mistake. She had caught its mistake while doing the job the system was supposed to support.

That distinction mattered during the next eight weeks. When she watched the screen, she sometimes lost the start of the next Spanish sentence. When she kept her eyes on the speaker, the transcript continued filling with words she could not review. Errors were easiest to find after a complete thought, but by then testimony had moved forward and an interruption could break a witness’s pace.

The printed page stayed in her work folder. She brought it to a meeting about the pilot and placed a finger beside the missing “ex.” She did not argue that every live transcript was unusable. She argued that a person interpreting cannot also provide continuous quality control without giving up attention somewhere else.

What the machine changed

A lost note or careless clerk could also produce a wrong word, but neither would create this particular job. The live system generated a second version of testimony during the hearing, displayed it with sentence-like punctuation, and made error review seem available at no added cost because a language worker was already present.

That was the change inside Elena’s work. Accuracy used to mean listening for the source speaker’s meaning and reproducing it in another language. Under the pilot, accuracy also meant checking what a machine believed it heard from her English, even though its failures were shaped by microphone placement, overlapping voices, unfamiliar names and the probability of one phrase following another.

The tool performed best under conditions a courtroom could not maintain. People turned away from microphones. Attorneys began speaking before witnesses had stopped. Some speakers lowered their voices around medical details or addresses.

Accents varied, and relationship terms arrived inside longer accounts rather than as isolated vocabulary.

Elena found uses for the transcript. During a break, it helped locate a disputed date without replaying a full audio segment. A staff member with hearing loss could follow portions of a routine hearing more easily. In clear audio, the text provided a rough map of what had been said.

She did not want it removed. She wanted its limits reflected in the work.

A rough transcript can be reviewed after a hearing, when an interpreter can compare it with audio and notes without missing new testimony. Live review is different. The screen competes with the speaker at the moment the interpreter has the least attention to spare, and the errors most likely to matter are often small words that a fluent sentence can hide: “not,” “former,” “before,” or the “ex” on Elena’s sheet.

The pilot also left privacy questions open for her. The display could contain names, medical information, family relationships and temporary addresses. Elena knew who could see the monitor in the courtroom. She was not given a clear account of how long draft text remained stored, whether it became part of other digital case materials, or who could review it after the hearing.

Those questions were separate from whether a proceeding was open to the public. Spoken testimony passes through a room and an official record follows defined practices. Machine-generated draft text can be copied, searched and retained, even when nobody has established that each line is accurate. In sensitive hearings, a wrong relationship or address can travel with the right names around it.

After the pilot, the courthouse kept the transcription tool for some proceedings. The expectation that interpreters monitor every line became less firm, though it did not vanish. Some staff still pointed to the screen when text looked odd. Elena would check when the pace allowed it.

She no longer treated the display as another speaker requiring constant attention.

The sheet with “wife” remained useful because it showed both sides of the tool on one line. The sentence was readable. Most of it was right. The smallest missing part carried the fact that mattered.

Questions people ask

Can a live

AI transcript replace a court interpreter?

In the experience behind this composite, no. The system transcribed spoken English but did not carry meaning between languages, manage incomplete phrases or resolve context from a witness’s account. It also produced fluent-looking errors. The interpreter remained responsible for the spoken interpretation while the transcript served as a rough secondary record.

Why would speech recognition drop a word like “ex”?

Short sounds can be masked by overlap, low volume or room noise. A speech-recognition system ranks likely word sequences from audio patterns, and “my wife” may score as a cleaner or more common sequence than “my ex-wife” when the first syllable is weak. The displayed sentence can still look complete.

Is monitoring the transcript part of interpreting?

The interpreters represented here experienced it as a separate attention task. They could review a line or flag an obvious error, but continuous monitoring pulled attention from incoming testimony. The courthouse’s pilot showed that adding a screen did not add more listening capacity or create time for accuracy review.

What happened to the incorrect transcript line?

Elena raised the difference during the proceeding, and the relationship term was clarified. The live transcript was described as a backup rather than the official record, but she was not given a full account of how draft text was retained. Her own copy stayed in a folder, with a circle around “wife” and a handwritten “ex.”