One entryMerged PR· crunch.stumason.dev· PHP
PR #22 merged: Revert ASR_THREADS=7 — it regressed transcription speed
additions
+0
deletions
-4
files changed
1
Measured ASR_THREADS=7 on the live box: consistently slower, not faster — 62-73s wall for 44.2s of audio (~0.6-0.8× realtime) across 3 clean warm runs, vs the ~2.7× baseline. The box is a 16-core ARM (CAX41) with shared vCPUs; CTranslate2's default already uses all cores for the small large-v3-turbo int8 model, so explicitly pinning 7 just capped it lower. Reverting to the default restores the fast path.
Real speedup levers for CPU transcription (not done here — accuracy tradeoffs, your call): beam_size 5→1, vad_filter=true (skip silence), or a GPU box.
🤖 Generated with Claude Code