Skip to content

bench(asr): radxa (RK3588) SenseVoice results + Whisper blocker - #76

Open
suharvest wants to merge 1 commit into
mainfrom
bench/results-radxa
Open

suharvest wants to merge 1 commit into
mainfrom
bench/results-radxa

Conversation

@suharvest

Copy link
Copy Markdown
Collaborator

Summary

  • SenseVoice zh (AISHELL-1, 20 items) on radxa/RK3588 via rk3588-sensevoice: concurrency=1 clean (mean CER 5.99%, p50 1292ms, RTF p50 0.216); concurrency 2/4/8 correctly rejected with too_many_sessions — traced and confirmed empirically that the session ceiling is hardcoded (RKASRBackend.concurrency_capability(max_concurrent=1)), not adjustable via OVS_MAX_CONCURRENT_SESSIONS.
  • Whisper (en) not benchable on this device today: rk.whisper is registered on main (commit 7cc9dd55) but absent from every image cached or pullable on radxa (rk-qwen3asr-opt-20260610, rk-20260903.10, rk-20260803b); the one image that validated RK3588 Whisper was never pushed to the registry. Root-cause detailed in bench/asr_bench/results/radxa.md.
  • Root-caused (not patched blindly) two blockers on the DISPATCH-default June image before switching to the Sept 3 image: missing sentencepiece, and SenseVoice reporting no STREAMING capability (predates the offline-pseudo-streaming shim).
  • Device cleaned up: no asrbench-* containers/volumes left, disk back to pre-task baseline, pre-existing edge_retail_console stack and other containers untouched.

Test plan

  • results/radxa-sensevoice-zh.json / .md — raw + summary bench output
  • results/radxa-sensevoice-zh-resource.csv — CPU/mem/NPU-sysfs samples during the sweep
  • results/radxa.md — full writeup with EVIDENCE section (docker ps/df before-after, log excerpts, git merge-base checks)
  • Whisper — blocked, not run (see root-cause in radxa.md)

🤖 Generated with Claude Code

https://claude.ai/code/session_016aHr1wmPGB5MLFoMe72Maf

…cution_policy test

SenseVoice zh (AISHELL-1 subset, 20 items) on rk3588-sensevoice via
seeed-local-voice:rk-20260903.10 (already cached): c=1 clean (20/20,
mean CER 5.99%, p50 1292ms, RTF p50 0.216); c=2/4/8 correctly hard-fail
with too_many_sessions. Traced the ceiling to
RKASRBackend.concurrency_capability() (hardcoded max_concurrent=1, NPU
exclusive device) and confirmed empirically that
OVS_MAX_CONCURRENT_SESSIONS=4 is clamped back to 1 with a warning — not
a config knob, no second sweep to run.

Root-caused two blockers on the DISPATCH-suggested rk-qwen3asr-opt-20260610
image before switching to rk-20260903.10: missing sentencepiece dependency,
and SenseVoice reporting OFFLINE-only (no STREAMING) capability because
that image predates the offline-pseudo-streaming shim added in commit
2a3cabb.

Whisper: not benchable on this device today. rk.whisper was registered in
commit 7cc9dd5 (on main) but is absent from every pullable/cached image
tried (rk-qwen3asr-opt-20260610, rk-20260903.10, rk-20260803b) — traced via
git merge-base that 7cc9dd5 postdates rk-20260903.10's base commit. The
one image that did validate RK3588 Whisper (openvoicestream:rk-whisper-rel,
per bench/perf/whisper/results_backend/PROVENANCE.md) was never pushed to
the registry. Building a new image is out of scope for this bench pass.

Cleanup verified: docker ps -a / df -h back to pre-task baseline; no
asrbench-* container or volume left on the device.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016aHr1wmPGB5MLFoMe72Maf
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant