Java: Text to Speech engines overview

Question

Dear Community!

I'm now in a search for Java Text to Speech (TTS) framework. During my investigations I've found several JSAPI1.0-(partially)-compatible frameworks listed on JSAPI Implementations page, as well as a pair of Java TTS frameworks which do not appear to follow JSAPI spec (Mary, Say-It-Now). I've also noted that currently no reference implementation exists for JSAPI.

Brief tests I've done for FreeTTS (first one listed in JSAPI impls page) show that it is far from reading simple and obvious words (examples: ABC, blackboard). Other tests are currently in progress.

And here goes the question (6, actually):

1. Which of Java-based TTS frameworks have you used?
2. Which ones, by your opinion, are capable of reading the largest wordbase?
3. What about their voice quality?
4. What about their performance?
5. Which non-Java frameworks with Java bindings are there on the scene?
6. Which of them would you recommend?

Thank you in advance for your comments and suggestions.

Answer 1

I've used Mary before and I was very impressed with the quality of the voices. Unfortunately, I haven't used any of the other ones.

Answer 2

I've actually had pretty good luck with FreeTTS

Answer 3

Thanks a lot everyone, the trick is in FreeTTS source. Briefly: if being run as 'java -jar freetts.jar some-more-args-here', it spells lesser words than when being executed in a manner of bin/Server.jar and bin/Client.jar.

Thanks to everyone once again!

Answer 4

I've used AT&T Natural Voices which provides JSAPI and MS SAPI hooks. It provides excellent quality voices, a good "general" speech dictionary, many controls over pronunciation, and multiple languages. It's a little pricey, but works very well.

I used it to read important sensor telemetry to drivers in a mobile sensor application. We had no complaints about the voice quality. It had about 75% out-of-the-box accuracy with scientific terms and a much higher (maybe 90%+) with normal dialogue. We got it up to about 99+% accuracy by using markups (most errors were on scientific terms with unusual phoneme combinations).

It was a bit hard on the processor (we were running on a Pentium-III equivalent machine and it was pushing 50%-75% peak CPU). This uses a native speech engine (Windows, Linux, and Mac compatible) with a Java interface.

There's a huge variety of voices and languages...

Answer 5

I used FreeTTS but had a major problem getting the MBrola voices to run on My MacbookPro. I did get MBrola voices to run on Windows (painfully) and Linux. I've had no luck loading any other voice packages on FreeTTS which is a shame because the supplied voices are horrible IMO. Outside of that I had a little success with Cloudgarden as well but that only runs on Windows AFAIK. I'd be interested to hear others successes/failures with Voice engines as this type of work is particular challenging. I'm also toying a bit with Sphinx4. I just pulled down JVXML (which appears to be based on Sphinx4) last night but could not get it to run for some strange reason.

Answer 6

I've contributed to mary. I feel it has potential if someone smarter than me separated the HMM voices out of the core (those voices don't need large data sets and sound ok). I'm also trying to do a event system to freetts to send events when it says a word. I've had success, but it is broken in linux now. (probably because of a timer bug).