nobody132/masr

Built on an end-to-end gated convolutional neural network (inspired by Facebook's Wav2letter but using GLU activation instead of ReLU/HardTanh for faster convergence), MASR trains on the AISHELL-1 dataset (150 hours, 4000+ Chinese characters) and achieves 14% character error rate on test sets. The architecture supports external language models to further reduce error rates to 8%, though performance remains limited compared to industrial systems trained on significantly larger datasets with domain-specific language models.

1,964 stars. No commits in the last 6 months.