rolczynski/Automatic-Speech-Recognition

Implements modular ASR pipelines with CTC loss and greedy decoding, achieving 6.71% WER on LibriSpeech dev-clean without external language models. Built on TensorFlow 2.0 eager execution, it decouples feature extraction, neural network models, and decoders as independent components—allowing swapping of sophisticated feature extraction, data augmentation strategies, or language model decoders. Supports distributed training via TensorFlow Strategy and mixed-precision policies for optimized inference.

223 stars. No commits in the last 6 months.