WeNet

WeNet positions itself as a production-first end-to-end speech recognition toolkit, and its defining trick is the U2 family of Conformer models that serve streaming and non-streaming recognition from a single network, letting one trained model power both live captioning and batch transcription. The Python side covers data pipelines and training recipes for standard corpora, while the runtime side ships libtorch-based deployments for x86 servers and Android, a gRPC service, and ONNX and TensorRT export paths for optimized serving. Pretrained checkpoints include WenetSpeech Chinese models alongside support for Whisper, Paraformer, and FireRed models within the same tooling. Basic usage installs with pip from the repository, but the intended workflow, training or fine-tuning a model and then compiling the runtime, assumes GPU training capacity and some systems comfort. WeNet is heavily deployed in Chinese industry and is the framework underneath Rev's Reverb ASR release. Everything is Apache-2.0 licensed, the repository has about 5,200 stars, and documentation lives at wenet-e2e.github.io.