Orze-ASR-3Way

Orze-ASR-3Way is a deterministic English ASR consensus system selected by Orze for the Open ASR Leaderboard. It uses bosonai/Qwen3-ASR-1.7B-hf-orze as its anchor and two independent voters: HojoAI/Hojo-ASR-V1 and OpenMOSS-Team/MOSS-Transcribe-preview-2B.

The anchor transcript is retained unless both voters agree on the same aligned word edit. A substitution or deletion is accepted only when both voters agree; an insertion is accepted only when both voters produce the same insertion at the same anchor boundary. The official English normalizer is applied before word alignment. The same rule and decoding parameters are used on every dataset.

Immutable components

Role	Checkpoint	Revision
Anchor	bosonai/Qwen3-ASR-1.7B-hf-orze	bce834cee3d106a9f0ec1ba98e96d995f13f59de
Voter 1	HojoAI/Hojo-ASR-V1	a22c381896ce5f4b70038982bfb888fe9969ed99
Voter 2	OpenMOSS-Team/MOSS-Transcribe-preview-2B	c4b3988677df13c14e79d9db59f356ed761db366

The reproducible Docker evaluator is hosted at erik-at-boson/open-asr-leaderboard-orze-ensemble. It runs the checkpoints sequentially and includes the sum of all three inference times in RTFx.

Open ASR Leaderboard results

These results use the official evaluator and normalizer at commit d1e99b25524814332d6868a5645e568670834cfb.

Dataset	WER (%)
AMI Cleaned	6.93
Earnings22 Cleaned AA (chunked)	5.24
GigaSpeech Cleaned	6.68
LibriSpeech test-clean	1.02
LibriSpeech test-other	2.41
SPGISpeech	1.58
VoxPopuli Cleaned AA	2.78
Mean	3.81

Measured aggregate model throughput is 29.66 RTFx, calculated from the sum of all three component inference times. The reproducible evaluator also records the small consensus-alignment overhead when it generates fresh manifests.

Held-out conversational proxy

The same fixed consensus rule was evaluated on three disjoint deterministic selections of 200 segments from the evaluation-only AppTek multi-accent call-center corpus. These selections were not used for training. Lower WER is better.

Selection	Anchor overall	Consensus overall	Anchor accent macro	Consensus accent macro
1	10.23	9.88	11.274	10.786
2	7.24	6.99	7.024	6.762
3	9.46	9.15	9.216	8.954

Both individual voters were worse than the anchor on all three selections; the unanimity gate nevertheless improved the anchor consistently. No private Open ASR Leaderboard data was accessed.

The fixed rule was also checked on a deterministic, balanced 300-utterance MInDS-14 holdout spanning en-AU, en-GB, and en-US. Overall WER improved from 31.24 to 31.04, and locale-macro WER improved from 31.32 to 31.13. The consensus changed 87 normalized transcripts: 38 improved, 34 worsened, and 15 were error-neutral, for 10 fewer word errors in aggregate. Hojo and MOSS individually scored 52.37 and 39.18 WER on the same selection.

Scope and limitations

This is a 9.64B-parameter composite system (2.04B Orze Qwen, 5.18B Hojo, and 2.42B MOSS) optimized for English short-form ASR. It trades throughput and memory footprint for accuracy. The models run sequentially, so peak inference memory is the maximum of the components rather than their sum. See each component card for its training disclosure and limitations; the Orze anchor card contains the fine-tuning disclosure for the only checkpoint trained as part of this work.

License

The consensus code and all three component checkpoints are Apache-2.0.