Ishiki-ASR-0.5 (ishiki-labs/ishiki-asr-0.5)
Built on NVIDIA Nemotron
ishiki-labs/ishiki-asr-0.5 (v0.5) is an automatic speech recognition (ASR) model developed by Ishiki Labs Inc., built on nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B.
License & Attribution (
OpenMDW-1.1): This model is a derivative work ofnvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4Band is governed by the NVIDIA Open Model License Agreement (OpenMDW-1.1). Original base model copyright (c) 2026 NVIDIA Corporation & Affiliates. Derivative model modifications copyright (c) 2026 Ishiki Labs Inc. SeeLICENSEandTHIRD_PARTY_NOTICES.mdfor full terms.
Open ASR Leaderboard Benchmark Results (English Short-Form, v0.5)
| Benchmark Dataset | WER (%) |
|---|---|
| AMI-Cleaned | 5.96 |
| Earnings22-Cleaned-AA-chunked | 4.11 |
| GigaSpeech-Cleaned | 6.15 |
| SPGISpeech | 1.54 |
| URGENT 2024 | 8.69 |
| URGENT 2024 (Clean) | 3.63 |
| Voice Arena Monsoon (en-IN) | 2.62 |
| VoxPopuli-AA-Cleaned | 1.43 |
| LibriSpeech (test-clean) | 0.83 |
| LibriSpeech (test-other) | 1.65 |
| Mean WER (8 Main Short-Form Sets) | 4.27 |
| Mean WER (10 Public Leaderboard Sets) | 3.66 |
Evaluation Usage
python ishiki_labs/run_eval.py \
--model_id="ishiki-labs/ishiki-asr-0.5" \
--checkpoint_path="ishiki-labs/ishiki-asr-0.5" \
--dataset_path="hf-audio/open-asr-leaderboard" \
--dataset="ami_cleaned" \
--split="test" \
--batch_size=16
Training, Testing, and Evaluation Dataset Disclosure
ishiki-labs/ishiki-asr-0.5 is built on nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B. No benchmark evaluation test sets from the Open ASR Leaderboard were used during optimization. Full base model training dataset disclosure is provided at: