Listen, Attend and Spell: Beam Search Pruning
Intermediate
The problem, from first principles
LAS decodes with beam search, keeping the best few partial transcripts at each step rather than the single best greedy choice. Pruning to the top scores is the step that keeps the search tractable.
From theory to code
Implement top_k_beams(scores, k), the indices of the k best hypotheses.
Constraints
- Ties keep the earlier index.
Hints
Hint 1
Sort the indices by descending score with a stable sort, then keep the first k.
Keywords
research-paperssequence-modelsspeechattention
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.