FunASR

FunASR began as Alibaba DAMO Academy's Paraformer serving stack and has grown into the broadest open speech recognition toolkit anyone currently maintains. The toolkit source is MIT. The model weights are not automatically MIT, and that distinction is the thing to settle before you deploy: Fun-ASR-Nano-2512 and Fun-ASR-MLT-Nano-2512, both 800M parameters, are Apache 2.0, while several other checkpoints ship under the FunASR Model Open Source License Agreement. The README says as much, quietly, and the model cards are the authority.

Fun-ASR-Nano-2512 handles Chinese, English and Japanese along with seven Chinese dialect groups (Wu, Cantonese, Min, Hakka, Gan, Xiang and Jin) and 26 regional accents, reporting 1.80% CER on open Chinese benchmarks and 1.76% WER on Librispeech-clean. Fun-ASR-MLT-Nano covers 31 languages with an East and Southeast Asian bias. Beyond those, the same API gives you SenseVoiceSmall at 234M for transcription with emotion and audio-event tagging, the Paraformer streaming variants for low-latency work and Whisper-large-v3 alongside its turbo build if you want a familiar baseline in the same harness. The pipeline pieces that usually cost the most integration time are included rather than left as an exercise: voice activity detection, punctuation restoration, timestamp alignment and speaker diarization.

Deployment options are unusually good for a project this close to research. There is a vLLM path for batch throughput, a llama.cpp GGUF build that quantizes to roughly 484MB and runs on CPU, a WebSocket streaming service and OpenAI-compatible endpoints, so existing Whisper client code often works with a base-URL change. Because everything is self-hosted, none of the account friction around Alibaba's hosted speech APIs applies: no mainland phone number, no business licence and no ICP filing. The trade is documentation. The English README is maintained but the deployment guides, runtime tuning notes and the issue backlog are Mandarin-first, and the project moves fast enough that community English write-ups go stale within a few months of publication.