Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

RNNT modified_adaptive_expansion_search is too slow.

See original GitHub issue

Hi, I try to use modified_adaptive_expansion_search with beam_size 5 to recognize one wav, which costs 52.9s, while ALSD costs 0.87S. why the speed is so slow.

Issue Analytics

State:
Created a year ago
Comments:5

Top GitHub Comments

2reactions

b-flocommented, Apr 21, 2022

Hi,

Can you give information? recipe, architecture, units, etc?

Nevertheless, I’m assuming you are modeling BPE or use a big vocab size? If so, the issue is known for ESPnet1 and due to select_k_expansions(...), see here

mAES: Fixed select_top_k_candidates loop taking a long time with a big vocab size. It also improves decoding time in general.

I’ll fix it in ESPnet1 when I have time, for the moment please either use the changes from PR #4032 or use another time-synchronous algorithm.

0reactions

b-flocommented, Apr 24, 2022

It shouldn’t make a difference in practice but yes, it should be recombine_hyps + sorted instead of sorted + recombine_hyps (pruning should be applied at the end anyway). Will change that in #4032.