Fun-ASR

Why we included this project

Chinese dialects and accented speech are where many ASR models fall apart, so this family from Tongyi Lab is worth knowing if your audio leans that way. The Nano checkpoint is trained on tens of millions of hours of Chinese, English, Japanese, and dialect audio, while the 800M MLT-Nano variant extends coverage to 31 languages. Both integrate with FunASR for inference and serving, with vLLM handling batch transcription and a llama.cpp path for GGUF models on edge devices. Runnable examples cover streaming transcription and speaker diarization, so the jump from clone to working pipeline is short. If you want a self-hosted alternative to closed ASR APIs, especially for Chinese-heavy workloads, this is a practical place to start.