Submission withdrawn September 19, 2026 (UTC). Open ASR PR #221 is closed. The leaderboard submission has been withdrawn; the checkpoint, code, research results, and exposure disclosures remain available as historical research artifacts.
This repository preserves the original R15-0100 Orukeet checkpoint originally proposed as an earlier checkpoint for Open ASR submission #221. The artifact is unchanged from the archived September 6, 2026 model. It has a FastConformer encoder and TDT decoder derived from NVIDIA Parakeet TDT 0.6B v3, with multilingual/accent adaptation and fitted temporal Gabor kernels materialized in the native weights.
The NeMo file is orukeet-r15-0100.nemo (2,509,342,720 bytes), SHA-256 4295a6d820a40b99786331d1c7a6b6c328916c8329b23d39415b0649a5d42811. Weights use CC BY-SA 4.0. Loading requires no custom model code:
from huggingface_hub import hf_hub_download
from nemo.collections.asr.models import ASRModel
path = hf_hub_download("oruk/orukeet-r15-0100", "orukeet-r15-0100.nemo", revision="073489c0619cd7939e327bebfc6c5d4ace4b69bf")
model = ASRModel.restore_from(path)
Training and evaluation history
R15-0100 precedes the later FT-4035 continuation whose training partitions were selected using an evaluation that included Monsoon English. It also precedes the later LibriSpeech test-other direct fine-tuning used to produce the released r3 checkpoint. Neither later stage is present in these weights.
Earlier evaluation and selection exposure remains. The original adaptation recipe included LibriSpeech test-other among validation/evaluation inputs, and the recovery selection leading to R15-0100 used a regression suite containing all 2,620 LibriSpeech test-clean recordings. Earlier VoxPopuli evaluation also occurred; its overlap with the current cleaned partition has not been established. The initial training recipe configured checkpoint monitoring, but the audit has not established whether test-other affected the particular fixed-step/EMA export inherited by R15-0100. Upstream pretraining overlap is not independently certified.
This model is an earlier checkpoint before the two later stages described above, not a claim of no benchmark exposure. The submission is withdrawn; acceptance is not claimed. The r3 public-eight WER of 4.49% and its corresponding placement do not apply to this checkpoint. Fresh results for R15-0100 are reported below. See the full disclosure and checkpoint provenance.
Public English Open ASR evaluation
Evaluated September 19, 2026 (UTC), with all 74,443 eligible recordings from the eight public English datasets. These are self-reported measurements; the submission is withdrawn, and no official acceptance or private-set placement is claimed. The archived checkpoint was selected by lineage before this run, without a new checkpoint or decoding sweep.
Profile: NVIDIA H200, NeMo 2.7.2, Torch 2.8.0, BF16, batch 128, greedy-batch decoding, max_symbols=10, one data-loader worker, original audio boundaries and official reference filtering. Timing covers one full transcription pass, bracketed by CUDA synchronization, after up to four warmup batches. Batch 128 follows the existing Parakeet profile; exhaustive throughput tuning is not claimed.
Pinned evaluator and reproduction instructions; complete results and output hashes.
| Dataset |
WER (%) |
H200 RTFx |
| ami_cleaned |
9.69 |
3625.50 |
| earnings22_cleaned_aa_chunked |
6.50 |
2980.66 |
| gigaspeech_cleaned |
7.95 |
5474.17 |
| librispeech_test_clean |
1.49 |
4076.97 |
| librispeech_test_other |
3.11 |
4144.22 |
| spgispeech |
3.39 |
6700.20 |
| voicearena_monsoon_en_in |
3.98 |
4862.19 |
| voxpopuli_cleaned_aa |
3.07 |
2620.26 |
| Public-eight aggregate |
4.90 |
5689.99 |
R15-0100 predates the later direct LibriSpeech test-other fine-tune and Monsoon-informed FT-4035 continuation. Earlier LibriSpeech test-clean regression selection and LibriSpeech test-other validation/evaluation remain. The earlier val_wer monitor followed FLEURS in inspected logs, with test-other reported separately; the influence of test-other evaluation on the fixed-step/EMA export is unresolved. Earlier VoxPopuli evaluation also occurred, with overlap against the current cleaned partition unverified. Inherited pretraining overlap is not independently certified. This is not a fully unseen or held-out checkpoint; the submission is withdrawn and eligibility is not claimed.
Monsoon has no registered Hub metric task. All eight results are retained above; no task ID is invented.
RTFx aggregate is total audio divided by total measured inference time, not an arithmetic mean of per-dataset RTFx.
