ASRT is a deep-learning-based Chinese speech recognition system that converts spoken Mandarin audio into text using convolutional neural networks, long short-term memory networks, attention mechanisms, and CTC loss.
The system addresses the challenge of recognizing Mandarin speech by combining CNN and LSTM architectures with attention mechanisms and connectionist temporal classification. It processes audio input through these neural network layers to produce character-level transcriptions. The project supports multiple Chinese speech datasets including THCHS30, ST-CMDS, Primewords, AISHELL-1, AiDataTang200, and MagicData, with configuration files allowing users to select which datasets to use for training.
Adoption requires substantial computational resources: training demands a GPU with at least 11GB of memory, 16GB of system RAM, and a 4-core processor, along with 500GB of storage. The tool runs on Linux for both training and inference, or Windows for inference only, with Python 3.9 through 3.11 and TensorFlow 2.5 through 2.11. Users should expect to download and prepare datasets before beginning training. The project provides multiple model variants, currently supporting models 24, 25, 251, and 251bn. It includes an HTTP API server for deploying the trained model as a service, making it suitable for applications requiring programmatic access to speech recognition capabilities. The system is designed for developers with the infrastructure to support GPU-accelerated deep learning workflows targeting Mandarin speech recognition.
The project maintains active issue tracking and documentation, with the author requesting detailed system specifications and complete error screenshots when problems are reported. A community QQ group and WeChat contact are provided for user support and discussion. The README emphasizes consulting project documentation and FAQ resources before submitting questions to avoid duplicate inquiries.
