Quiz

If you need to watch the videos again, or refer to the readings whilst answering the questions, that’s allowed.

An IFF target cost function can use any or all of the linguistic features provided by the front end. Using more features seems like a good idea, because it will allow the system to retrieve more closely-matched candidates from the database. What are the downsides of using a large number of features in an IFF target cost function?

The feature space grows with every new feature added - for example, adding a feature that can take 5 possible values multiples the size of the feature space (i.e., the number of unique feature value combinations) by 5. This increases sparsity and reduces the chances of finding exact, or even close, matches from the database. Assuming a simple weighted-sum formulation of the cost function, each new feature will also need a weight to be set, making tuning of the set of weights harder.

Consider two competing sequences of candidates being considered for synthesising a particular target sequence. Sequence A has quite well-matched linguistic features, so would have a low total target cost using an IFF target cost function. Sequence B has poorly-matched linguistic features, but has a very low target cost using an ASF target cost function. Do you think sequence B will sound better than sequence A?

An ASF target cost function relies on accurate predictions of acoustic properties for the target, and a perceptually-correlated way of measuring their similarity to the available candidates. If both of those things are true, sequence B is likely to sound better.

Following from the previous question, consider a system that uses the linguistic features from the front end to make very accurate predictions of F0 and phone duration. These predicted properties are then used in an ASF target cost function. Do you think this system would generate speech with perfect prosody?

As Taylor (16.2.4, with references back to chapter 9) points out, there is more to prosody than F0, so this approach is not guaranteed to produce better prosody than a IFF system using the same linguistic features.