Haas-ja-v1.0 has 615.8 million parameters and was trained on about 46,000 hours of data. The company reports the best strict CER among evaluated open models on asr-ja-bench v1.1, which contains 18,318 utterances and 23.38 hours of Japanese speech.

Estimated infrastructure cost is $0.003–$0.004 per audio hour, compared with about $0.30 for GPT-Transcribe. This is an infrastructure estimate under high-utilization assumptions, not an API price; operating costs may rise with light use. The model requires Linux and an NVIDIA GPU.

Why it matters

It offers an openly licensed option for teams that want to self-host Japanese transcription, with a company-reported low inference-cost estimate. The benchmark and cost claims are not presented as independently verified.

GUIDES

Something to try

As a hypothetical trial, check the license and hardware requirements, then evaluate a short Japanese recording relevant to your work and compare it with your current transcription using the same metric.