Skip to content

Our measurement · Speech recognition

Dictation Accuracy on Long Recordings, Tested

Do speech recognition engines transcribe several minutes of continuous speech as accurately as the same speech handed over in short pieces?

6 results, measured by usUpdated October 25, 2026Free to reuse under CC BY 4.0

Download CSVJSONEvery transcript and timing

Every data point

Each row is one published figure for one group, in percent. The row opens its fact page with the exact words of the source; the source opens the original.

Source
Apple SpeechTranscriber, in one recordingApple SpeechTranscriberAppleOne recording of about 4 minutes8440English2.01 to 2.67Our measurement
Parakeet TDT v3, in one recordingParakeet TDT v3NVIDIAOne recording of about 4 minutes8440English2.36 to 3.06Our measurement
Whisper large-v3-turbo, in one recordingWhisper large-v3-turboOpenAIOne recording of about 4 minutes8440English2.54 to 3.27Our measurement
Apple SpeechTranscriber, in piecesApple SpeechTranscriberAppleThe same audio in pieces of 10 to 20 seconds8440English2.01 to 2.65Our measurement
Whisper large-v3-turbo, in piecesWhisper large-v3-turboOpenAIThe same audio in pieces of 10 to 20 seconds8440English2.08 to 2.63Our measurement
Parakeet TDT v3, in piecesParakeet TDT v3NVIDIAThe same audio in pieces of 10 to 20 seconds8440English2.1 to 2.83Our measurement

6 rows. Open a row for the exact words of its source.

Calculated from the rows

Every statistic is calculated by code from the rows above. Open one to see the exact values it summarises.

How this dataset is built

What counts as a row
In every chapter of the test-clean split of LibriSpeech-Long (Google DeepMind, 2024, from LibriSpeech's audiobook recordings), its first example of at least 3 minutes: 84 readings of 3 to 4 minutes by all 40 of the split's speakers, 5.45 hours of clean read English. Each reading is transcribed twice by each engine: as one recording, and cut at pauses into pieces of 10 to 20 seconds whose text is joined.
What does not
Readings under 3 minutes (a chapter's tail), the split's noisier recordings (test-other), other languages, people dictating their own words, streaming as an app transcribes while you speak, and cloud engines, for now.
How rows are found and checked
Each engine transcribed every reading on a Mac with English given. The pieces end at the quietest tenth of a second between 10 and 20 seconds after each piece starts, so they split the speech at pauses, as someone dictating in short bursts would. Word error rate is computed against LibriSpeech's own transcript with the same normalization as our main benchmark, and the same alignment counts the words an engine left out and the longest run of them. Speakers, not readings, are the unit of the margins of error: intervals come from resampling speakers with all their readings, the two conditions are compared on the same readings, and a difference counts as measured only when its interval, allowing for the test's comparisons, excludes zero. Every file's checksum, cut point, reference and transcript are in the full record.
What the data cannot tell you
Audiobook narrators reading clean text in a quiet room are easier to transcribe than a person composing an email aloud, and their pauses are regular. The test hands each engine a finished recording; an app that transcribes while you speak, or cuts the audio its own way, may do better or worse than either condition. LibriSpeech writes numbers as words, so an engine that writes digits is charged a little in both conditions alike: compare each engine with itself across the two conditions for what length does, and read the engine against engine comparisons with that in mind.

The whole method, from search to calculation, is on the research data page.

Cite and reuse

You may copy, share and adapt this dataset for any purpose, including commercially, if you credit it with a link. Quoted passages remain their authors' words: cite the original source for those.

Yap (2026). Speech recognition accuracy on long dictation: one 4-minute recording against the same speech in short pieces [Data set]. https://yapapp.cc/data/asr-long-dictation

License: CC BY 4.0.

More datasets

Read the overview: Speech Recognition: What It Is and How Well It Works