Quiz

This module has a two-part quiz – some flashcards, then a speed reading exercise.

In HMM-based speech synthesis, why might separate regression trees be employed for F0 and for the spectral envelope?

Think about how a regression tree (a type of CART) performs regression: by asking a sequence of questions about the predictors (i.e., the linguistic context). The order of the questions matters, with those that are most predictive appearing closer to the root of the tree. Since different aspects of the context will be most useful for predicting F0 (e.g., position-in-phrase) than for predicting the spectral envelope (e.g., local phonetic context), separate trees will give better results.

What is the relationship between the regression tree in the "HMM speech synthesis, viewed as regression" video and the tree used to cluster models in the "HMM speech synthesis, described as context-dependent modelling" video?

They are exactly the same thing, it is only the way they were described that differs. You can interpret the tree as either performing regression, or as clustering models together. Whichever way you prefer to understand it, the tree is predicting the parameters of the HMMs, from the linguistic context.

Given that the regression tree is performing the key task of getting from linguistic context to speech (model) parameters, what is the point of the HMM?

The regression tree is not a sequence model: it simply operates on one data point at a time. We need an external timekeeping mechanism. During synthesis, this provides duration information: deciding when to move forward in the linguistic context sequence. During training, it is responsible for aligning the linguistic context sequence with the speech feature sequence (for an HMM, we can use the Viterbi algorithm or Expectation-Maximisation).

Speed reading exercise

Then answer the following questions – you are allowed to refer back to the paper, but you must locate the answer as quickly as possible.