NL-2610.009Ongoing · Planned[nl.SA]
Whisper, size by size: speed against accuracy on real recordings
The Nook team
Opened 5 October 2026 · 1 min read
Nook's Transcribe and Translate speech Nooklets use Whisper, which comes in sizes from tiny to large. Bigger is more accurate and slower, but by how much, on the hardware people own? We will transcribe one set of real recordings with every size, on an RTX 4060 and on the processor alone, and report speed against word error rate, so the choice of size can rest on numbers.
- Stage
- Planned
- Comments
- RTX 4060 8 GB and the processor; whisper.cpp, CUDA and CPU builds.
- Subjects
- Speech and audio (nl.SA); Performance (nl.PF)
- Cite as
NL-2610.009· usenook.ai/labs/whisper-sizes.html
Full text
The question
For a one-hour meeting, which Whisper should you use? The tiny model is fast and makes mistakes; the large one is careful and slow. Somewhere between them is the size that is good enough for the job and quick enough to wait for, and it is probably different on a graphics card and on a laptop's processor.
The plan
- Gather real recordings with human-checked transcripts: a clear talk, a meeting with several voices, a phone call, and speech in a language other than English.
- Transcribe every recording with every Whisper size, with whisper.cpp's CUDA build on the RTX 4060 and its processor build.
- Report speed as minutes of audio per minute of work, the word error rate against the checked transcript, and peak memory.
- Publish the recordings we are allowed to share, and the scripts.
So far
Planned; no runs yet. Recordings with checked transcripts that we may publish are the hard part. If you have some to share, write to hello@usenook.ai.