AI, Human and Stenographic Transcription: How Accuracy Compares

Three Ways to Produce a Hearing Transcript

Stenography. A stenographic reporter writes what is said in shorthand on a stenotype machine, and translation software converts the strokes into English text. The National Court Reporters Association (NCRA) describes that text as a feed that can be read, searched and archived as the proceedings run, and lists courts, depositions and arbitrations among the settings where reporters work. It states that reporters reach speeds upwards of 280 words per minute.

Transcription from a recording. Transcripts can also be typed after the event from an audio recording. In the courts and tribunals of England and Wales, for example, a party requests a transcript, court staff locate the audio recording, and the court’s authorised transcription company provides the transcript; that company may also have provided a stenographer or court logger to make a record of the proceedings.

AI speech-to-text. Speech recognition engines generate a transcript from the audio as people speak. At TERES this is the first layer of the service, not the finished record: trained human transcribers review the AI output in real time during the hearing.

What Accuracy Depends On

A single accuracy figure only means something alongside the conditions it was measured under. The TERES service level agreement, for example, applies its accuracy commitments to audio that meets minimum standards: clear speech without significant background noise, recorded at a sample rate of at least 16 kHz.

In a hearing, several advocates, witnesses and tribunal members may speak in turn, with varied accents and specialist legal terminology. TERES trains its system on varied accents and legal terminology, and relies on its human review layer to get speaker identification and context right.

How the TERES Human Review Layer Works

TERES runs transcription as a managed service. AI speech-to-text engines generate the transcript live during the hearing, and trained human transcribers review it in real time, so the transcript the tribunal and counsel rely on has been checked by a person rather than left as raw machine output.

The transcript is verbatim as standard, with speaker identification and timestamps, and a cleaned-up version is available on request. The real-time transcript is available during the hearing, and the final version follows within a few hours of the hearing closing.

The Accuracy Figures in the TERES SLA

The TERES service level agreement sets two accuracy figures, and they describe different stages of the same transcript.

TERES measures delivery and accuracy separately. The SLA’s on-time delivery rate is the share of transcripts delivered within the stated delivery time and is not a measure of accuracy. The 99.9% uptime figure in the SLA is the monthly availability commitment for the DASH platform core, not a transcript accuracy figure.

• AI transcription: at least 98% accurate for clear audio
• After review by TERES human transcribers: 99.5% accurate

Comparing Accuracy Figures Like for Like

Figures quoted for different methods are often measured in different ways. Before comparing providers or methods, it helps to ask:

• Does the figure describe raw output, or the transcript after a person has reviewed it?
• What audio conditions does the figure assume, and what happens when a hearing falls short of them?
• Is the figure a contractual commitment, a test threshold or a marketing claim, and where is it published?
• Does it cover speaker attribution as well as the words themselves?
• Is it an accuracy figure at all, or a delivery or uptime figure?