hirofumi0810/tensorflow_end2end_speech_recognition

Implements multiple encoder architectures (BLSTM, GRU, VGG variants) with diverse decoding strategies including CTC with beam search and attention mechanisms (Bahdanau, Luong, location-based). Supports joint CTC-Attention training, multi-GPU synchronous training, and multi-task learning with auxiliary CTC losses at arbitrary encoder layers. Handles multiple datasets (TIMIT, LibriSpeech, CSJ) across various output units (phonemes, characters, words, kanji) with preprocessing via a separate dedicated repository.

314 stars. No commits in the last 6 months.