WFST-based ASR decoding

For the Automatic Speech Recognition course, I implemented a decoder built on weighted finite-state transducers — the classical backbone of ASR systems before end-to-end neural approaches took over. The work involved WFST composition and optimisation, beam-pruned Viterbi decoding, explicit silence modelling, and integration with KenLM for n-gram language model scoring.

This was an exercise in understanding the engineering beneath the abstractions. Modern ASR systems hide this machinery inside neural architectures, but the problems it solves — efficient search over exponentially large hypothesis spaces, principled integration of acoustic and language model scores — haven’t gone away. They’ve just been reformulated.