masr

Why we included this project

MASR is a compact, end-to-end Mandarin speech recognition project built on a gated convolutional network, structurally similar to Facebook's 2016 Wav2letter work but using GLU activations rather than ReLU or HardTanh. The author trains on the AISHELL-1 corpus and publishes character error rates, including what an external language model buys you: test CER drops from about 14% to 8%. Pretrained weights close to the best run come with the repo, so you can try it without training from scratch. The README is also unusually honest about scale, noting that its 150 hours of audio is a small fraction of what commercial systems train on. Teams studying conv-based acoustic modeling, or just curious what a solo-maintained Chinese ASR project can realistically do, will find this a readable reference rather than a production drop-in.