zai-org/GLM-ASR

Built on a multimodal architecture compatible with Hugging Face Transformers and SGLang, GLM-ASR-Nano excels at low-volume and dialectal speech recognition—particularly Cantonese and regional Chinese variants—achieving 4.10% average error rate across challenging real-world benchmarks like meeting transcription. The model supports 17 languages through a unified chat-based interface, enabling flexible audio-to-text inference via standard APIs or direct model loading.

759 stars. Actively maintained with 1 commit in the last 30 days.