Reaction
Open speech model Hojo-ASR-Multi-V1 reportedly posts 3.54 average WER across five languages
A user says it pairs an acoustic encoder with a Qwen3-4B decoder and is available under Apache 2.0.
TLDR
A user describes Hojo-ASR-Multi-V1 as an open multilingual speech-recognition model with a reported 3.54 average word error rate across five languages. It combines an acoustic encoder, adapter and Qwen3-4B decoder, is available on Hugging Face under Apache 2.0, and supports local inference. The user suggests testing difficult audio conditions and measuring memory use, throughput and latency before judging its suitability for edge hardware.
Combined views
—
3 Sources, first seen ago
