v0.1.0

Initial release of Pruned CTC, a memory-efficient CTC loss implementation for large-vocabulary speech recognition.

Highlights

• Exact loss and first-order gradient equivalence under vocabulary reduction.
• Head-and-loss activation memory that no longer scales linearly with vocabulary size.
• Configurable finite-beam alignment pruning, which introduces a separate approximation.
• Support for float32, bfloat16, and arbitrary blank token IDs.
• Public API: pruned_ctc_loss.
• MIT license.

Installation

pip install pruned-ctc==0.1.0

Requires Python 3.10+, PyTorch 2.4+, and a separately installed k2 build matching your PyTorch/CUDA environment.

View on PyPI