Quiz

Here are some questions about the videos in this module. If you need to watch the videos again whilst answering the questions, that’s allowed.

Unit selection (and as we will see later, SPSS and neural methods) uses a finite-sized corpus of recorded speech but can synthesise the spoken form of an infinite variety of input texts. For unit selection, what is the key insight that makes us believe this is possible?

Unit selection assumes that there is not an infinite variety of spoken forms. More specifically, although there is an infinite variety of complete utterances, they can be constructed from smaller units which exhibit limited variability. The crucial question to be answered is: what is that smaller unit and how many types do we need?

Why cannot the target cost simply compare the speech waveform in a candidate with the corresponding waveform in the target?

Because target units do not have waveforms! They have linguistic features, provided by the front end. They might optionally have some acoustic properties - such as F0 and duration - that have been predicted from those linguistic features. But we do not go as far as actually predicting a complete waveform - that would take us into the realm of SPSS and neural methods, coming later in the course.

Are there any restrictions on what features or properties the join cost can take into account?

Provided the features or properties are available for each of the two candidates being considered for joining (concatenation), there are no limits. In other words, the join cost can be based on any candidate features available locally. Using non-local features (e.g., properties of other candidates) would complicate the search and potentially prevent the use of dynamic programming.