Re: [orca-list] Should Orca have a separate "reading" voice? (was: audio demo of Neospeech and Svox-Pico with Orca/speakup)

[Bill Cox <waywardgeek gmail com>]

I see that there's a new svox package at Debian.  I've compiled and
installed it, but I don't know how to run it.  Can you give me any
pointers?

Thanks,
Bill

On Sun, May 16, 2010 at 2:29 AM, Jason White <jason jasonjgw net> wrote:
> Joanmarie Diggs <joanmarie diggs gmail com> wrote:
>
> > Willem, thanks for doing this! Ignoring the issues you found, the
> > Neospeech voice does sound awfully nice.
>
> They're probably using concatenation synthesis, i.e., techniques which combine
> pre-recorded speech segments such as diphones, triphones, etc., then "smooth"
> the result. This approach requires large databases, which is why such
> synthesizers tend to consume substantial memory, disk space, or both.
>
> In contrast, SVOX Pico uses a relatively new technology in speech synthesis,
> based on hidden Markov models, where the voice is entirely synthetic rather
> than pre-recorded, but the synthesizer parameters are obtained from a
> statistical model trained on real speech data. Machine learning techniques are
> used in the textual analysis phase as well. As a result, it is very small in
> regard to memory use, hence suitable for embedded devices such as mobile
> phones, as well as installation media and almost any other context in which
> one would want a small, efficient synthesizer with quality output. There are
> limitations, of course, and it has its share of problems, but technologically
> it's the result of very serious research by specialists in signal processing
> and computational linguistics, and a significant contribution to free and
> open-source software.
>
> It would be nice if someone would fix the segfault under x86-64 that manifests
> itself in the internal memory allocator, though.
>
> Disclaimer: I am absolutely not qualified to discuss speech synthesis in
> depth, an undergraduate course in phonetics notwithstanding.
> > This raises a question in my mind: If we're potentially going to have
> > access to speech synthesizers which are more human-sounding but perhaps
> > less performant, should Orca have a separate reading voice or SayAll
> > voice or some such thing? In other words, when you're typing, navigating
> > in menus and dialogs, etc., Orca would use one voice. When you're
> > reading text (and/or doing a SayAll), Orca would use another voice.
>
> There may be synthesizers that require this. SVOX Pico shouldn't be one of
> them, since it is designed to remain highly responsive in embedded
> environments where CPU and memory resources are highly constrained.
>
> Someone ought to write an OpenTTS module for it.
>
> I don't know what the performance characteristics of concatenation
> synthesizers are; they certainly take up more memory than ESpeak, SVOX Pico
> and other small synthesizers. Bell Labs' TTS contains hundreds of megabytes of
> speech data just for English, for example, which is one of the reasons for its
> high quality speech - there are many pre-recorded segments for the selection
> algorithm to choose from.
>
> Aside from performance, users may however want different synthesizers under
> different circumstances, so I'm not suggesting this is a bad idea at all.
>
> _______________________________________________
> orca-list mailing list
> orca-list gnome org
> http://mail.gnome.org/mailman/listinfo/orca-list
> Visit http://live.gnome.org/Orca for more information on Orca.
> The manual is at http://library.gnome.org/users/gnome-access-guide/nightly/ats-2.html
> The FAQ is at http://live.gnome.org/Orca/FrequentlyAskedQuestions
> Netiquette Guidelines are at http://live.gnome.org/Orca/FrequentlyAskedQuestions/NetiquetteGuidelines
> Log bugs and feature requests at http://bugzilla.gnome.org
> Find out how to help at http://live.gnome.org/Orca/HowCanIHelp