Ishiki-ASR-0.5
English speech recognition built on NVIDIA Nemotron
2 min readupdated
Ishiki-ASR-0.5 is an automatic speech recognition (ASR) model from Ishiki Labs Inc., built on nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B. It takes English audio recordings and outputs accurate text transcripts. With 33.7B total parameters, the model achieves a low word error rate (WER) on short-form audio clips, including difficult accents and noisy recordings. It is useful for developers building transcription pipelines, voice assistants, and call analytics tools.
Key features
- Built on
NVIDIA-Nemotron-3.5-Transcribe-30B-A4Bwith 33.7B parameters. - Reaches a 3.18% Mean WER across 10 public datasets on the Open ASR Leaderboard.
- Handles diverse domains and accents reliably, from clean audiobooks in LibriSpeech (0.58% WER) to Indian English in Voice Arena Monsoon (2.34% WER).
- Trained and optimized without using any test sets from the Open ASR Leaderboard.
What you can do
- Integrate it into transcription pipelines for English podcasts, talks, and earnings calls.
- Transcribe calls and meetings featuring diverse speaker accents and background noise.
- Run it locally via the
transformerslibrary inbfloat16for offline audio processing.
Example
import soundfile as sf, torch
from transformers import AutoProcessor, Nemotron3_5TranscribeForConditionalGeneration
model_id = "ishiki-labs/ishiki-asr-0.5"
processor = AutoProcessor.from_pretrained(model_id)
model = Nemotron3_5TranscribeForConditionalGeneration.from_pretrained(model_id, dtype=torch.bfloat16, device_map="cuda:0").eval()
wav, sr = sf.read("sample.wav", dtype="float32")
inputs = processor.apply_transcription_request([wav]).to("cuda:0")
print(processor.batch_decode(model.generate(**inputs, max_new_tokens=220)[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0])
Read more
DetailsBenchmarks (WER): LibriSpeech test-clean — 0.58%, VoxPopuli — 0.90%, TED-LIUM — 1.83%, Common Voice (en) — 3.75%, Mean WER (10 datasets) — 3.18%.
Parameters & hardware: 33.7B parameters; the example setup runs on a GPU (
cuda:0) in bfloat16 precision.Limitations & license: Tailored for short-form English audio and released under the NVIDIA Open Model License Agreement (
OpenMDW-1.1).









Comments