These project ideas involve vocals.
Glossolalia vocals are vocals whose lyrics are written or improvised at the phonemic level (as opposed to the word level in an existing language).
Some glossolalia lyrics consist of words of a novel language: thus, the songwriter first invents the syntax and semantics of a new language, and then writes a song in the language. Elizabeth Fraser is believed to use this approach -- the Cocteau Twins song Carolyn's Fingers uses this technique, available as a legal live recording here (the studio recording is much better, however).
Other glossolalia lyrics exist only at the phonetic level, and thus are free of semantics. It appears that Lisa Gerrard's lyrics usually take this form. A legal download example of her style is available here; most of the Dead Can Dance albums have at least one track of this nature (some of the DCD tracks are a cappella).
As this FAQ entry indicates, Sigur Ros also uses phonetic-level glossolalia. Beth Gibbons (Portishead) has also investigated the technique, as she discusses in this (rare) interview.
Glossolalia vocals are interesting for concatenative synthesis for several reasons:
Several types of projects come to mind in this domain. One idea (suggested by Adrian Freed) is to design tools for composers to analyze existing glossolalia languages and to design new languages. The a cappella recordings of Lisa Gerrard may be good source material for an analysis project.
Another project idea is to analyze the artifact properties of vocal synthesis algorithms, and design generative algorithms for glossolalia languages and for glossolalia compositions that minimize artifacts.
Another project idea is to investigate real-time speech synthesis from a MIDI controller, by leaving phoneme selection to an online composition system.
In lecture, we discussed a children's choir product as an example of a successful phrase-based concatenative system. The children sang Latin text phrases, which could be automatically strung together. Given that most people do not understand spoken Latin, most listeners would not notice semantic lyrical issues in the spliced Latin, and would also be more tolerant of audio artifacts.
These nice properties won't work for phrase-based vocal synthesis aimed at popular music in the native language of the listeners. In class, we showed an example of a phrase-based pop music vocal product whose lyric selection seemed to be selected in an ad hoc way, without concern to the semantics and rhyme structure of concatenation.
An interesting project would be to consider phrase-based pop song vocal libraries in a principled way. Such a project would be driven both from data (such as Web lyric databases for a particular song style) and from rules about the structure of lyric and melodic phrases in the music form. Ideally, the project would include a small pilot test (recording a phrase library produced by the phrase generation system, and hand-concatenating it to produce songs).
A good introduction to the structure of musical lyrics in pop music is available in this five-part series: [1] [2] [3] [4] [5]. For vocal phrasing ideas, the book analyzed in this review may also be helpful (it is available in the music library on the stacks).
The articles linked off of this webpage provide a good sample of the machinery used for the statistical analysis of language. To learn more, see the the book described on this webpage. The webpage contains the Preface and Chapter 1 of the current edition, and draft versions of many of the other chapters for the upcoming edition. These chapters should be enough to let you know if computational linguistics has a role in your focus area for the project (in which case you probably want to read the paper version of the book).