ASH ASR workbench. Moonshine LibriSpeech scores and the asr.jezzick.net report.
  • Python 57.2%
  • HTML 42.8%
Find a file
2026-09-28 11:25:29 -05:00
config Score Moonshine Tiny Q8 on LibriSpeech test-clean and test-other. 2026-09-28 11:10:13 -05:00
reports Restore the comparison layout and label the single-model LibriSpeech limit. 2026-09-28 11:15:51 -05:00
scripts Score Moonshine Tiny Q8 on LibriSpeech test-clean and test-other. 2026-09-28 11:10:13 -05:00
HANDOFF.md Record the pause: one scored model, other runtimes later. 2026-09-28 11:25:29 -05:00
README.md Score Moonshine Tiny Q8 on LibriSpeech test-clean and test-other. 2026-09-28 11:10:13 -05:00

asr-workbench

ASH-local speech recognition bench. The Rowan inference-lab-benchmarks harness was the starting checkpoint for the result shape. The runs, audio, and scored report are on ASH.

What this scores

Moonshine Streaming Tiny Q8 (private-moonshine-asr) on the official LibriSpeech test sets:

Backend Set Corpus WER Edits / words RTF
CPU 127.0.0.1:18197 test-clean 4.84% 2,546 / 52,576 0.0067
GPU1 127.0.0.1:18198 test-clean 4.87% 2,560 / 52,576 0.0071
CPU test-other 12.94% 6,773 / 52,343 0.0070
GPU1 test-other 13.12% 6,870 / 52,343 0.0076

Corpus WER is total word edits divided by total reference words. Tokenizer: lowercase letters, digits, apostrophes. This is not the official SCTK scorer.

The public page is reports/index.html, served at asr.jezzick.net. It does not include LibriSpeech audio or any private recording.

Layout

  • scripts/librispeech_eval.py — transcribe a LibriSpeech subset and append jsonl
  • scripts/summarize_librispeech.py — corpus WER summary
  • reports/ — scored jsonl, summary JSON, and the public HTML
  • config/cpu-asr.json — the CPU audio.cpp server config used for the run

LibriSpeech itself stays on ASH at /mnt/models/librispeech-20260928 and is not in this repo.