Hojo-ASR-Multi-V1 achieves 3.54 WER across five languages
Summary
Hojo-ASR-Multi-V1, an open multilingual automatic speech recognition (ASR) model, has achieved an average word error rate (WER) of 3.54 across five European languages. This model combines an acoustic encoder with a Qwen3-4B decoder, enabling it to leverage a comprehensive language model during the decoding stage, resulting in enhanced speech recognition performance. The model is publicly available for developers to install and run locally via the hojo-asr Python package, and it supports various audio input formats. Released under the Apache 2.0 license, Hojo-ASR-Multi-V1 allows for academic, personal, and commercial use, highlighting its accessibility and relevance in advancing multilingual conversational speech recognition.
Analysis
Qwen3-4B: Qwen3-4B is a large language model from the Qwen series that functions as the decoder in specialized speech recognition systems. In this setup, it supplies semantic and contextual language knowledge to resolve ambiguities in encoded audio features after an adapter layer bridges the acoustic encoder output. This integration allows the overall model to draw on richer word-sequence modeling during transcription. Hugging Face: Hugging Face is a platform for hosting and distributing machine learning models and related resources. The Hojo-ASR-Multi-V1 model is hosted there, enabling direct access, inspection, and local execution through community tools and packages. It supports open licensing that facilitates both research and secondary development. Hojo-ASR-Multi-V1: Hojo-ASR-Multi-V1 is an open multilingual automatic speech recognition model built with an Encoder-Adapter-LLM architecture that incorporates a Qwen3-4B decoder for the final transcription stage. It targets conversational speech recognition across multiple languages and emphasizes local, inspectable inference. The model was recently made publicly available on Hugging Face under an open license, with support for various input formats via a dedicated Python package. Licensing: The model is released under the Apache 2.0 license to support academic, personal, and commercial secondary use. Architecture: The system separates acoustic encoding from language-based reasoning by routing adapted audio features into an LLM decoder for final transcript generation. Availability: Developers can install and run the model locally using the hojo-asr Python package with support for multiple input types including WAV files and byte streams.
Categories
aimachine_learningtech