Pruning in unit selection search procedure

[Georgia M]

Watching the video the “Search Procedure” I am trying to visualise the best unit sequence search as token passing and understand how pruning is implemented in unit selection synthesis.

While using Viterbi algorithm in speech recognition, we were keeping the single token with the highest probability every time more than one tokens met. On top of that, while implementing pruning we threw away locally ‘winner’ tokens if they were below some level. In this way, despite making the search faster, it was possible to prune the path that would have gone on to “win”, so error rates might increase.

In unit selection synthesis we are choosing the token bearing the lowest cost (sum of target cost of every unit in sequence + join cost between every pair of consecutive units in sequence).In addition, we implement pruning which could theoretically lead again to eliminating winner tokens.
I would like to know if implementing pruning in this case has any side-effects in the model’s performance, and how this could be evaluated. I am having a hard time trying to deduce this..

The speech recognition token passing video had been of great help and I am trying to make a similar visualisation to the unit selection search problem.

Thanks

[Simon]

We’ll cover this in the lecture.

[Author]

Posts