mravanelli/pytorch-kaldi

The toolkit bridges PyTorch's model flexibility with Kaldi's robust acoustic processing through a modular DNN/HMM architecture, supporting multi-stream inputs and complex neural topologies (MLPs, CNNs, RNNs, LSTMs, SincNet) configured via intuitive YAML files. Key capabilities include automatic feature chunking with context expansion, multi-GPU training, HPC cluster integration, and plug-in support for custom acoustic models without modifying core code. Built-in scheduling for dynamic hyperparameter adjustment (batch size, learning rate, dropout) during training enables sophisticated training regimens.

2,396 stars. No commits in the last 6 months.