Audio modelOpen weights

Orukeet

Oruk

Released: September 11, 2026
Data date: October 7, 2026

Model class: Multilingual speech-recognition model
Languages: 25
Access: Public weights for local use; no hosted inference provider listed

Orukeet is a multilingual speech-recognition model based on NVIDIA Parakeet TDT 0.6B v3. The released checkpoint has 627 million parameters and is listed for 25 languages.

Oruk replaces half of the temporal filters with fitted, frozen Gabor kernels. The weights use CC BY-SA 4.0 with NVIDIA Foundation attribution, while the code is MIT-licensed.

Specifications and access

Specification	Value and source
Model class	Multilingual speech-recognition modelSource
Access	Public weights for local use; no hosted inference provider listedSource
Input modalities	AudioSource
Output modalities	TextSource
Checkpoint	oruk/orukeetSource
Base model	NVIDIA Parakeet TDT 0.6B v3Source
Parameters	627 millionSource
Languages	25Source
License	Weights and fitted kernels: CC BY-SA 4.0 with NVIDIA Foundation attribution. Code: MIT.Source
Hosted inference	Not deployed by any inference provider, according to Hugging FaceSource

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

LibriSpeech test-clean CER (Show measurement, test conditions, and source)

Source value: 0.56
Score: 0.56
Metric: character error rate before compound alignment
Unit: %
Benchmark version: Exact r3 checkpoint; published per-partition scores
Category: audio
Direction: Lower is better
Sample: 2620
Source type: vendor-reported
Evaluator: Orukeet r3 paired recognition score ledger
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Exact Orukeet r3 NeMo checkpoint SHA-256 031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56; matched fresh decoding against stock Parakeet TDT v3 (3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d), greedy-batch TDT, FP32 weights and BF16 CUDA autocast; LibriSpeech test-clean, 2,620 clips.

Source Orukeet r3 paired recognition score ledger

LibriSpeech test-other CER (Show measurement, test conditions, and source)

Source value: 1.19
Score: 1.19
Metric: character error rate before compound alignment
Unit: %
Benchmark version: Exact r3 checkpoint; published per-partition scores
Category: audio
Direction: Lower is better
Sample: 2939
Source type: vendor-reported
Evaluator: Orukeet r3 paired recognition score ledger
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Exact Orukeet r3 NeMo checkpoint SHA-256 031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56; matched fresh decoding against stock Parakeet TDT v3 (3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d), greedy-batch TDT, FP32 weights and BF16 CUDA autocast; LibriSpeech test-other, 2,939 clips.
Context: Oruk states that all 2,939 test-other recordings were used for final adaptation and checkpoint selection. This is a reported development/selection score, not an independent holdout.

Source Orukeet r3 paired recognition score ledger

FLEURS Bulgarian CER (Show measurement, test conditions, and source)

Source value: 3.34
Score: 3.34
Metric: character error rate before compound alignment
Unit: %
Benchmark version: Exact r3 checkpoint; published per-partition scores
Category: audio
Direction: Lower is better
Sample: 658
Source type: vendor-reported
Evaluator: Orukeet r3 paired recognition score ledger
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Exact Orukeet r3 NeMo checkpoint SHA-256 031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56; matched fresh decoding against stock Parakeet TDT v3 (3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d), greedy-batch TDT, FP32 weights and BF16 CUDA autocast; FLEURS Bulgarian, 658 clips.

Source Orukeet r3 paired recognition score ledger

FLEURS Croatian CER (Show measurement, test conditions, and source)

Source value: 3.67
Score: 3.67
Metric: character error rate before compound alignment
Unit: %
Benchmark version: Exact r3 checkpoint; published per-partition scores
Category: audio
Direction: Lower is better
Sample: 914
Source type: vendor-reported
Evaluator: Orukeet r3 paired recognition score ledger
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Exact Orukeet r3 NeMo checkpoint SHA-256 031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56; matched fresh decoding against stock Parakeet TDT v3 (3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d), greedy-batch TDT, FP32 weights and BF16 CUDA autocast; FLEURS Croatian, 914 clips.

Source Orukeet r3 paired recognition score ledger

FLEURS Czech CER (Show measurement, test conditions, and source)

Source value: 2.67
Score: 2.67
Metric: character error rate before compound alignment
Unit: %
Benchmark version: Exact r3 checkpoint; published per-partition scores
Category: audio
Direction: Lower is better
Sample: 723
Source type: vendor-reported
Evaluator: Orukeet r3 paired recognition score ledger
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Exact Orukeet r3 NeMo checkpoint SHA-256 031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56; matched fresh decoding against stock Parakeet TDT v3 (3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d), greedy-batch TDT, FP32 weights and BF16 CUDA autocast; FLEURS Czech, 723 clips.

Source Orukeet r3 paired recognition score ledger

FLEURS Danish CER (Show measurement, test conditions, and source)

Source value: 5.31
Score: 5.31
Metric: character error rate before compound alignment
Unit: %
Benchmark version: Exact r3 checkpoint; published per-partition scores
Category: audio
Direction: Lower is better
Sample: 930
Source type: vendor-reported
Evaluator: Orukeet r3 paired recognition score ledger
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Exact Orukeet r3 NeMo checkpoint SHA-256 031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56; matched fresh decoding against stock Parakeet TDT v3 (3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d), greedy-batch TDT, FP32 weights and BF16 CUDA autocast; FLEURS Danish, 930 clips.

Source Orukeet r3 paired recognition score ledger

FLEURS Dutch CER (Show measurement, test conditions, and source)

Source value: 1.93
Score: 1.93
Metric: character error rate before compound alignment
Unit: %
Benchmark version: Exact r3 checkpoint; published per-partition scores
Category: audio
Direction: Lower is better
Sample: 364
Source type: vendor-reported
Evaluator: Orukeet r3 paired recognition score ledger
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Exact Orukeet r3 NeMo checkpoint SHA-256 031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56; matched fresh decoding against stock Parakeet TDT v3 (3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d), greedy-batch TDT, FP32 weights and BF16 CUDA autocast; FLEURS Dutch, 364 clips.

Source Orukeet r3 paired recognition score ledger

FLEURS English CER (Show measurement, test conditions, and source)

Source value: 1.77
Score: 1.77
Metric: character error rate before compound alignment
Unit: %
Benchmark version: Exact r3 checkpoint; published per-partition scores
Category: audio
Direction: Lower is better
Sample: 647
Source type: vendor-reported
Evaluator: Orukeet r3 paired recognition score ledger
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Exact Orukeet r3 NeMo checkpoint SHA-256 031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56; matched fresh decoding against stock Parakeet TDT v3 (3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d), greedy-batch TDT, FP32 weights and BF16 CUDA autocast; FLEURS English, 647 clips.

Source Orukeet r3 paired recognition score ledger

FLEURS Estonian CER (Show measurement, test conditions, and source)

Source value: 3.39
Score: 3.39
Metric: character error rate before compound alignment
Unit: %
Benchmark version: Exact r3 checkpoint; published per-partition scores
Category: audio
Direction: Lower is better
Sample: 893
Source type: vendor-reported
Evaluator: Orukeet r3 paired recognition score ledger
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Exact Orukeet r3 NeMo checkpoint SHA-256 031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56; matched fresh decoding against stock Parakeet TDT v3 (3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d), greedy-batch TDT, FP32 weights and BF16 CUDA autocast; FLEURS Estonian, 893 clips.

Source Orukeet r3 paired recognition score ledger

FLEURS Finnish CER (Show measurement, test conditions, and source)

Source value: 2.16
Score: 2.16
Metric: character error rate before compound alignment
Unit: %
Benchmark version: Exact r3 checkpoint; published per-partition scores
Category: audio
Direction: Lower is better
Sample: 918
Source type: vendor-reported
Evaluator: Orukeet r3 paired recognition score ledger
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Exact Orukeet r3 NeMo checkpoint SHA-256 031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56; matched fresh decoding against stock Parakeet TDT v3 (3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d), greedy-batch TDT, FP32 weights and BF16 CUDA autocast; FLEURS Finnish, 918 clips.

Source Orukeet r3 paired recognition score ledger

FLEURS French CER (Show measurement, test conditions, and source)

Source value: 1.7
Score: 1.7
Metric: character error rate before compound alignment
Unit: %
Benchmark version: Exact r3 checkpoint; published per-partition scores
Category: audio
Direction: Lower is better
Sample: 676
Source type: vendor-reported
Evaluator: Orukeet r3 paired recognition score ledger
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Exact Orukeet r3 NeMo checkpoint SHA-256 031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56; matched fresh decoding against stock Parakeet TDT v3 (3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d), greedy-batch TDT, FP32 weights and BF16 CUDA autocast; FLEURS French, 676 clips.

Source Orukeet r3 paired recognition score ledger

FLEURS German CER (Show measurement, test conditions, and source)

Source value: 1.52
Score: 1.52
Metric: character error rate before compound alignment
Unit: %
Benchmark version: Exact r3 checkpoint; published per-partition scores
Category: audio
Direction: Lower is better
Sample: 862
Source type: vendor-reported
Evaluator: Orukeet r3 paired recognition score ledger
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Exact Orukeet r3 NeMo checkpoint SHA-256 031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56; matched fresh decoding against stock Parakeet TDT v3 (3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d), greedy-batch TDT, FP32 weights and BF16 CUDA autocast; FLEURS German, 862 clips.

Source Orukeet r3 paired recognition score ledger

Page 1 of 18

Sources and data date

Type	Evidence and data date
Research status	Research date October 7, 2026. 7 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not found in the inspected sources	independent-exact-checkpoint-replication Original wording No external ASR replication of the exactly hashed r3 checkpoint was found in the sources read. The large WER/CER table comes from the model team.
Unresolved	test-other-holdout Original wording The model card says the r3 checkpoint was adapted on all 2,939 LibriSpeech test-other recordings and selected on that same split. Its score is not an independent test.
Unresolved	accent-domain-overlap Original wording Oruk says 6,118 accent/domain recordings entered preceding adaptation, but does not identify overlap with the reported 12,006 test recordings.
Unresolved	local-runtime-comparability Original wording The six-language checks use a small, disclosed FLEURS validation sample. Runtime timings have different included/excluded work and are not comparable across runtime rows.
Additional source	Oruk Orukeet release and evaluation report (retrieved October 7, 2026) · Oruk Orukeet release and evaluation report · Editorial description reviewed October 7, 2026
Additional source	Oruk Orukeet model card and license (retrieved October 7, 2026) · Oruk Orukeet model card and license · Editorial description reviewed October 7, 2026
Additional source	Orukeet r3 paired recognition score ledger (retrieved October 7, 2026)
Additional source	Oruk local transcription guide (retrieved October 7, 2026)
Additional source	Orukeet OpenWhispr application benchmark (retrieved October 7, 2026)
Additional source	Orukeet OpenWhispr application benchmark score JSON (retrieved October 7, 2026)
Additional source	Model metadata (retrieved October 7, 2026)