AliceBench
AliceBench v1, one number for how fast a computer transcribes with pktito, and how to run it.
What it is
AliceBench v1 times pktito transcribing LibriVox’s public-domain recording of Alice’s Adventures in Wonderland: twelve chapters, 2:58:51 in all, with the English model (parakeet-tdt-0.6b-v2). The result is the audio’s length over the time it took, as times real time.
Results
| Machine | × real time | Whole book | pktito | Backend | Threads |
|---|---|---|---|---|---|
| AMD Ryzen 9 9950X | 212.6 | 0:50 | 0.8.5 | avx2-vnni | 16 |
| NVIDIA DGX Spark (GB10) | 113.0 | 1:35 | 0.8.5 | neon-dp | 20 |
| Mac mini, M4 Pro | 106.1 | 1:41 | 0.8.4 | neon-dp | 12 |
| Raspberry Pi 5 | 14.4 | 12:25 | 0.8.5 | neon-dp | 4 |
Run October 23, 2026; every run matched all 12 reference transcripts. The Mac mini was in use at the time.
Run it
On macOS or Linux, from a clone of pktito:
scripts/alicebench.sh [DIR]
On Windows:
powershell -ExecutionPolicy Bypass -File scripts\alicebench.ps1 [DIR]
It downloads the chapters (200 MB) and the English model the first time, then prints each chapter’s speed and the total.
- DIR defaults to
~/.cache/alicebench-v1, or%LOCALAPPDATA%\alicebench-v1on Windows. PKTITO_CLInames thepktito-clito time; otherwise the one on the PATH, or the Mac app’s.
The full results are saved in DIR as result.json.
Why FLAC
The chapters are 16 kHz mono FLAC, made from LibriVox’s MP3s, because each system decodes an MP3 slightly differently: macOS’s and ffmpeg’s versions of chapter 11 differ in 5 of its 1,937 words. A FLAC decodes the same everywhere, so every computer transcribes the same audio.
“All 12 transcripts are the reference’s”
Every correct build of pktito writes the same transcript of the same audio, byte for byte, on any machine. So the script also checks each chapter’s transcript against the reference. That last line means this build computes the right answer, so its speed counts; a mismatch makes the script exit 1.