A TensorFlow 1.x time capsule of end-to-end speech recognition

What it does

This is an end-to-end speech recognition toolkit built in TensorFlow 1.x. It trains acoustic models—RNNs, LSTMs, GRUs, bidirectional variants, DeepSpeech2, and even a Capsule Network—to map audio features to character or phoneme sequences using CTC decoding. The project includes preprocessing pipelines for TIMIT, LibriSpeech, and WSJ, plus evaluation scripts that map TIMIT’s 61 phonemes down to 39 for scoring.

The interesting bit

The author publishes raw, unflattering LibriSpeech predictions—“wavraly” for waverley, “stragling bilagor” for straggling village—which is more honest than most 2017 ASR repos. That transparency makes it a useful fossil for seeing how far CTC-based models have come.

Key highlights

• Implements a laundry list of pre-Transformer architectures: BRNN, BLSTM, BGRU, Deep Residual Networks, Seq2Seq with attention, Layer Normalization RNN, and Capsule Networks.
• Supports TIMIT, LibriSpeech, and WSJ corpora with hardcoded MFCC extraction (39 coefficients) and phoneme mapping.
• Includes a basic n-gram language modeling module and CTC decoding.
• Runs on CPU or GPU, though it is locked to Python 3.5 and the TensorFlow 1.x ecosystem.

Caveats

• Stuck in the TensorFlow 1.x era; the last roadmap item from 2019 promises “TF1.12 soon,” which gives a sense of maintenance velocity.
• LibriSpeech output without a language model is visibly garbled, and the author notes the results used only “casual tuning because time it limited.”
• Depends on scikit.audiolab and system-level libsndfile, which are increasingly brittle in modern environments.

Verdict

Worth a look if you are studying the evolution of ASR or need a reference implementation of classical CTC architectures. Skip it if you want a modern, production-ready recognizer—this predates Whisper by half a decade.