Synthetic voices revisited

[Fidcal]

I recently took another look at what synthetic voices were available since last I looked a few years ago. In particular for use as text to speech. There has been a little progress but not a lot. There are more voices available now and the very best are only slightly better than a few years ago. There is still a trace of warble and the lesser voices are rather poor in this respect. OK if you just want some text read out to you but not suitable for producing game voices - which was my own main interest.

 

I could find almost no progress in dramatic control such as emotional content: anger, shouting, whispering, tenderness, humour. The control codes I found were extremely limited.

 

Pronounciation is only moderate. To some extent one can fool the voices into different pronounciations but it was neither easy nor adequate.

 

However, some use might have been made for neutral statements except for one blocker: highly restrictive distribution rights of output. I emailed my provider who gave me distribution prices of all the main voice companies and they were ridiculously high even for free, non-commerical useage. Note that I am talking about recordings of output made from their voices, not the original voice data. They were actually more expensive than hiring an actor. The worst case was quoted as "starting at $15,000..."

 

One interesting voice producer now only offers a web-based service. Effectively you pay a rather high monthly fee and submit your text for conversion to speech. Not sure how that would work because in my tests some years ago with software one can spend quite some time experimenting before you get it just right. Perhaps they 'watermark' the output while testing then only charge you for your final unwatermarked output.

 

Just today I'm reading about Salero - a combined effort to synthesis better animation and speech more easily. This is rather far ahead for my use but it somehow led me to the Vocaloid singing voices software. I was curious. Clearly singing requires some dramatic control of the synthesised voice. Maybe it could be used for speech which is less demanding than singing.

 

I didn't spend to much time but the website was rather obscure about obtaining its software and seemed only to offer an update. I did find http://www.zero-g.co.uk/index.cfm?articleid=802 however, from where you can download demos. I only tried one and it is immediately obvious it is more realistic than any of the speaking synthetic voices I have heard. What is more, they openly stated the output can be freely distributed (so long as you are not using say copyright lyrics or music of course.) If they can do this then why not the speaking voice companies? I would be happy to pay reasonable money for decent text to speech and distribution.