Selecting a speaker

We have found that choosing the right speaker to record, is actually as important as all the the other processes we describe. Some people just have better voices that are better for synthesis than others. In general people with, clearer, more consistent voices are better than others but unfortunately its not as clear as that. Professional speakers are in general better for synthesis that non-professional. Though not all professional voices work, and many non-professional speakers give good results.

In general you are looking for clear speakers, who don't mumble and don't have any speech impediments. It helps if they are aware of speech technology, i.e. have some vague idea of what a phoneme is. A consistent deliver is important. As different parts of speech from different parts of the recorded database are going to be extracted and put together you what the speech to be as consistent as possible. This is usually the quality that professional speakers have (or any one used to public speaking). Also note most people can't actually talk for long periods without practice. Lectures/Teachers are actually much more used to this than students, though this ability can be learned quite easily.

Note choosing the right speaker, if its important to you, can be a big project. For example, an experiment done at AT&T to select good speakers for synthesis involved recording fair sized databases for a number of professional speakers (20 or so) and building simple synthesis example from their voice and submitting these to a large number of human listeners to get them to evaluate quality [syrdal??]. Of course most of us don't have the resources to do searches like that but it is worth taking a little time to think of the best speaker before investing the considerable time in takes in building a speaker.

Note that trying to capture a particular voice is still somewhat difficult. You will always capture some of that persons voice but its unlikely a synthesizer built from recordings of a person will always sound just like that person. However you should note that voice you think are distinctive may be so because of lots variation. For example Homer Simpson's voice is distinctive but it would be difficult to built a synthesizer from. The Comic Book Guy (also from the Simpsons) also has a very distinctive voice but is much less varied prosodically than Homer's and hence it is likely to be easier to build a synthesizer from his voice. Likewise, Patrick Stewart's voice should be easier than Jim Carey's.

However as it is usually the case that you just have to take any speaker you have willing to do it (often yourself), there are still things you should do that will help the quality. It is best is recording is done in the same session, as it is difficult to set up the same recording environment (even when you are very careful). We recommend recording some time in the morning (not immediately you get up), and if you must re-record do so at the same time of day. Avoid recording when the speaker has a cold, or a hangover as it can be difficult to recreate that state if multiple sessions are required.