k2-fsa/k2

Health v2 · maintenance only

Adoption not part of the score

1352 stars · 237 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

k2 is a C++/CUDA library with Python bindings that implements differentiable Finite State Automaton (FSA) and Finite State Transducer (FST) algorithms integrated with autograd-based toolkits like PyTorch. It supports batched FST processing on CPU and CUDA, and is primarily used for automatic speech recognition tasks such as CTC loss, LF-MMI loss, and decoding with lattice rescoring.

Use cases

• compute CTC loss with FST-based decoding graphs
• train ASR models with LF-MMI objective in PyTorch
• perform fast pruned composition of dense FSA from neural network log-probs
• rescore lattices during multi-pass speech recognition decoding
• run differentiable FSA/FST algorithms on GPU with autograd support
• process batches of FSAs/FSTs in parallel on CUDA

When to choose

• you need differentiable FSA/FST operations inside PyTorch training loops
• you are building ASR systems with CTC or LF-MMI training objectives
• you need GPU-accelerated, batched FST decoding or lattice rescoring
• you want an extensible alternative to Kaldi's FST tooling

When to avoid

• you only need classical, non-differentiable FST tools like OpenFst for offline text processing
• your project uses TensorFlow (support not yet available)
• you need a pure-CPU solution without CUDA and lack pre-compiled wheel support for your platform
• you are not working with speech recognition or graph-based decoding

Facets

library · maturity active

machine-learning deep-learning speech-recognition math gpu-computing speech-processing machine-learning gpu-computing python cpp windows fsa fst autograd pytorch asr ctc lf-mmi ragged-tensor decoding lattice-rescoring natural-language-processing cuda linux macos

Member repositories

For agents

markdown · JSON · MCP: product_card(name="k2-fsa/k2")