Can Whisper transcribe long audio accurately?

Mostly, but length costs it something that its rivals pay less of. We ran Whisper large-v3-turbo through whisper.cpp on 84 English readings of 3 to 4 minutes. As one recording it had a word error rate of 2.91 percent word error rate ¹. The same audio cut at pauses into pieces of 10 to 20 seconds gave 2.35 percent word error rate ². That 0.56 percentage points of word error rate ³ is a measured difference, not chance.

In pieces, Whisper matched the other engines

On short pieces, Whisper was level with Apple SpeechTranscriber and Parakeet TDT v3. Its gap to Apple was within chance ⁴ and so was its gap to Parakeet ⁵. If you split audio before Whisper sees it, this test gives no reason to prefer another engine on accuracy.

One recording made it lose passages

Whisper left out a passage of 10 or more words in a row in 4 of 84 readings ⁶ when given the long recording, and in 0 of 84 readings ⁷ when given pieces. A dropped passage is the failure to watch for, since the text that remains reads normally.

Why a fixed window matters, and how timestamps carry it

Whisper reads audio in a fixed window of well under a minute. For anything longer, the software around the model has to move that window forward. In whisper.cpp the model predicts timestamps, and the program uses the last timestamp it trusts to decide where the next window begins. If a timestamp is wrong, or the model stops early inside a window, the next window can start after speech that was never written down. That is a plausible route to the missing passages we counted, although our test counts the symptom and does not diagnose the cause.

Cutting at pauses removes the problem from Whisper's side. Each piece fits inside one window, begins at a natural break and carries no earlier text for the model to be misled by.

Whisper paid more for length than Apple did

Compared by what one recording instead of pieces cost each engine, Whisper's cost was clearly larger than Apple's and not clearly different from Parakeet's:

Only the Whisper and Apple comparison is firm. Parakeet's own cost was 0.25 percentage points of word error rate ¹¹ and Apple's was 0.01 percentage points of word error rate ¹² (not distinguishable from chance). On the long recording, Apple's error rate was lower than Whisper's ¹³, while Parakeet and Whisper were level ¹⁴.

What this does and does not show

The readings are clean, read audiobook English from 40 speakers, run on a Mac with an M2 Pro. Your own dictation, other languages and noisy rooms will give different numbers (see our noise robustness data). We tested one model and one runtime, so other Whisper sizes, other front ends and their own splitting settings may behave differently. We did not test streaming or cloud engines. Read more about the measure in our guide to speech recognition accuracy.

What to do with long audio in Whisper

• Split at silences into pieces of 10 to 20 seconds, then join the text. That is the setup that gave no lost passages here.
• If your tool gives you only one long file, scan the transcript for jumps in the story, not only for spelling errors.
• If length is your main need and you want to compare engines, our market overview lists tools, and the on-device benchmark compares engines on other measures.

For workflow help, see how to transcribe a long audio file and how to dictate a book or long document on iPhone. The full dataset has every row, and our overview of speech recognition covers the basics.