Summary

• In a team of two students, we developed a speech-recognition system using weighted finite-state transducers (WFSTs) and a Viterbi decoder.
• Our goal was to reduce the word error rate and increase the computational efficiency of the algorithm.
• We did so by conducting various experiments on the system
  • Tuning transition probabilities, self-loop probabilities, and final probabilities.
  • Testing different WFSTs, based on uni-gram or n-gram word occurrence probabilities and adding optional silences between words.
  • Enhancing the Viterbi decoder by pruning the search-tree with different strategies.
  • Improving the efficiency of the decoder by using a tree-structured lexicon with language model look-ahead.
• We used the library openfst_python to model the WFSTs in this project and performed our experiments in a Jupyter notebook.

Main learnings

• I learned how I can communicate my ideas clearly to my teammate with a different background (computational linguistics). This was an interesting experience, since in most other projects, all team partners study computer science.
• Additionally, I now understand clearly how automatic speech recognition can work without neural networks. Of course, most current systems simply rely on machine learning.