philipperemy/tensorflow-ctc-speech-recognition

Uses LSTM networks with CTC loss to decode speech directly to text, trained and evaluated on the VCTK Corpus with configurable batch sizes and network architectures. Extracts audio features via librosa and python_speech_features, then feeds spectrograms through recurrent layers followed by CTC decoding to handle variable-length audio-text alignment without explicit frame-level annotations. Demonstrates end-to-end training on single-speaker subsets, showing reasonable generalization despite limited data through techniques like random silence truncation for realistic validation.

131 stars. No commits in the last 6 months.