Keep the recording beside the transcript

For an original newsroom workflow, keep the source recording beside the transcript. Review names, numbers, technical terms, and quotations against the audio before publication.

A readable sentence is not evidence that the speaker said it.

Speech recognition does not identify every speaker

Transcription and speaker identification are different tasks. Do not assume a speech recognizer can reliably assign every sentence to a person without an additional method and review.

What the system does

Whisper is an encoder-decoder trained to predict text from audio, described in OpenAI’s paper and released in several sizes. Larger sizes are slower and, on OpenAI’s reported evals, more accurate.

Accuracy here means the words match the audio more often, under their test sets. It does not mean the content is true.

A clear recording of a false sentence will transcribe as that false sentence.

The model also has translation behavior into English, depending on how you call it. A translation is a second task.

Errors in translation are not transcription errors. Say which task you ran.

Names, accents, and overlapping speech are the usual failure sites. A benchmark on clean read speech will overstate a meeting.

What to store

Store the model size, the language option, and a way to hear the span you are quoting. Do not store only the text and then treat it as minutes.

If a word is high stakes, a person should hear it. Automatic timestamps help that review.

They are not a substitute for it.

What to record before you rely on it

Record the Whisper size and whether you used a local file or an API that wraps it. Keep one example of a name it misspelled.

That example stops a team from calling the transcript a recording.

Keep the model id or the paper citation, the date you read it, and the inputs you used beside that note, so a later change can be compared with the same test.

When the result is a failure, store the artifact and change the design. Do not relax the check until the model happens to pass.

Name a person who owns that failure.

Common questions

Does a transcript prove what was said?

It is evidence about the audio, only as good as the model and the recording. For a dispute, keep the audio.

The text is a derived artifact.

Can I use Whisper output as training data without listening?

You will be training on the model’s mistakes. If the downstream use cares about names or numbers, sample the audio.

Do not assume a famous model’s transcript is clean.

Sources & further reading

  1. Robust Speech Recognition via Large-Scale Weak Supervision ↗
How this story was made

Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.

Our editorial standards
Back to all stories