ASRT_SpeechRecognition

Why we included this project

ASRT gives you a full Mandarin speech-to-text stack you can actually train yourself, not just a demo. It pairs a deep convolutional network with CTC decoding to turn audio up to 16 seconds long into pinyin, then passes that through a statistical language model to produce Chinese text. The repo bundles the whole workflow: training scripts that run on public datasets such as THCHS30, AIShell, and ST-CMDS, a downloadable trained model, and HTTP and gRPC servers with client SDKs for Python, Go, Java, and Windows. That makes it useful both for standing up a real transcription API and for seeing how a CNN plus CTC system fits together end to end. Plan on a decent NVIDIA GPU with roughly 11 GB of memory for training; CPU-only inference works on Linux and Windows.