What accuracy should I expect in practice?

On clear, close-miked single-speaker audio, modern systems are usually in the high nineties. Noisy multi-speaker recordings can fall well below that, and the drop is driven by conditions rather than by the choice of system.

How is accuracy actually measured?

Usually as 100% minus the word error rate against a human reference transcript, which means the reference and the normalisation rules both affect the result.

Can I improve accuracy without re-recording?

Supplying a vocabulary hint list of names and terms helps. Selecting the cleanest available audio track usually helps more.

Does a longer recording reduce accuracy?

Length itself does not. What correlates with length is the chance of conditions changing — a speaker moving away from the microphone, a new participant joining, background noise starting.

Definition

What Is Transcription Accuracy?

VideoText editorial team. Updated September 13, 2026.

Direct definition

Transcription accuracy is how closely a transcript matches what was said, under a stated definition of “match.” Researchers usually report word error rate. Clients may mean names, numbers, and style-guide compliance instead.

The sections below explain how transcription accuracy is used in transcription and subtitle work, what people often mix up, and which VideoText workflow applies when the next step is a file or review pass.

Key takeaways

Ask for the metric, the audio condition, and the sample.
Names and numbers can be “wrong” even when WER looks low.
A product accuracy figure is only useful when it names the dataset, sample, and metric — not as a standalone percentage.

How accuracy is judged

Research: align to a reference and compute WER or CER. Delivery work: a reviewer checks speakers, proper nouns, and the requested verbatim level.

Why condition matters

The Whisper paper’s Table 2 is the clearest public illustration: the same model’s WER ranges from a few percent on clean read speech to the mid-twenties or worse on noisy multi-speaker sets.

Why one accuracy figure is not enough

Accuracy is a function of the audio as much as the system. Clean single-speaker studio audio, a two-person video call, and a six-person meeting in a reverberant room will produce very different numbers from the same recogniser.

A published figure is therefore a statement about a test set. It predicts your results only to the extent that your material resembles that test set.

What moves the number most

Microphone distance and placement dominate. A lapel or headset microphone close to the speaker typically outperforms every downstream adjustment you can make afterwards.

After that: overlapping speech, background noise, strong accents relative to the training data, and domain vocabulary such as drug names, legal terms or product codes.

Frequently asked questions

What accuracy should I expect in practice?: On clear, close-miked single-speaker audio, modern systems are usually in the high nineties. Noisy multi-speaker recordings can fall well below that, and the drop is driven by conditions rather than by the choice of system.
How is accuracy actually measured?: Usually as 100% minus the word error rate against a human reference transcript, which means the reference and the normalisation rules both affect the result.
Can I improve accuracy without re-recording?: Supplying a vocabulary hint list of names and terms helps. Selecting the cleanest available audio track usually helps more.
Does a longer recording reduce accuracy?: Length itself does not. What correlates with length is the chance of conditions changing — a speaker moving away from the microphone, a new participant joining, background noise starting.

Related terms

Sources

OpenAI: Robust Speech Recognition via Large-Scale Weak Supervision