Add speakrs head-to-head harness and measured DER results - #89
Merged
Conversation
Wire speakrs into the cross-engine benchmark (single der.py scorer), add a thin speakrs-rttm CLI helper, refresh polyvoice runner variants for the v2+VBx default, and record the locked H2H protocol plus an interim 4-file matched-set DER summary.
10-file VoxConverse-test under der.py: speakrs-coreml 13.74% vs polyvoice 17.72% no-collar micro; gap is confusion-dominated. polyvoice remains faster on cold CLI (~53× vs ~23× CoreML on M1 Pro).
Process a directory of WAVs with a single loaded pipeline and --hyp-dir output, skipping cached RTTMs. Enables full-split CoreML runs in tens of minutes instead of cold-starting per file.
Warm batch on VoxConverse-test (M1 Pro, 144× RTFx) reproduces speakrs published accuracy. Gap vs polyvoice remains ~4 pp, confusion-dominated.
Matched der.py scoring on full VoxConverse-test. Gap is 4.14 pp no-collar micro and almost entirely confusion (8.04 vs 3.63).
Record speakrs CoreML 11.08% vs polyvoice 15.22% under the same der.py scorer (full 232, collar 0). Gap is confusion-dominated. Update COMPETITORS with speakrs as the primary Rust peer threat.
Same der.py scorer on Mix-Headset. Gap is 5.97 pp no-collar; speakrs gets speaker count roughly right (11/16 exact) while polyvoice collapses (2/16). Documents both corpus H2H tables in BENCHMARKS.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
speakrs_runner, warm-batchspeakrs-rttmhelper,score_h2h.py).benchmarks/der.pyon Apple M1 Pro:docs/BENCHMARKS.mdanddocs/COMPETITORS.md(speakrs as primary Rust peer).benchmarks/results/speakrs-h2h-2026-08-03/(PROTOCOL / VERDICT / JSON scores).Why
speakrs is the closest peer (Rust + ONNX/CoreML, community-1 style). Their published ~11% Vox DER is reproduced on our scorer. The residual is speaker confusion/count, not RTF.
Test plan
benchmarks/tools/speakrs-rttm(--features coremlon macOS)Notes
speakrscheckout by default (../../../../speakrsfrom the tool crate); seebenchmarks/tools/speakrs-rttm/README.md.