Labelling databases with HMMs

HMM that has many hundreds of states for each model. He shows that one can reduce die number of states by an arbitrary amount, allowing one to scale the size of a synthesis system in a principled manner. 

In all types of data-driven synthesis, we not only require data, but also require the data to be labelled in some way. As a minimum, this normally means that we require the words and phones, but any feature that we require for model building or unit selection must be provided by some means. In this section, we concentrate on one such way of carrying out this labelling, which is based on the HMM principles introduced above. 

The use of automatic labelling algorithms is normally justified in terms of saving time [5], [279], [384], In the past, we had small databases, which could be labelled by hand, but, since die databases we use today are larger, and because sometimes we wish to label die database very quickly, we require an automatic system. This argument basically says diat automatic labelling is therefore a matter of convenience if we had an infinite supply of hand labellers we wouldn t have a problem. A second justification is that automatic systems can outperform human labellers in terms of accuracy and consistency, so convenience alone is not the only justification automatic systems are in fact better. in our experience this certainly appears to be true. 

---

As the state to frame alignment is a by-product of recognition with HMMs, they can be used to label and segment a speech database. 

While many possible approaches to statistical synthesis are possible, most work has focused on using hidden Markov models (HMMs). This along with the unit selection techniques of the next chapter are termed third generation techniques. This chapter gives a full introduction to these and explains how they can be used in synthesis. In addition we also show how these can be used to automatically label speech databases, which finds use in many areas of speech technology, including unit selection synthesis. Finally, we introduce some other statistical synthesis techniques. 

---