Suggestions for Text-To-Speech (TTS) from Org sources?

[Jens Lechtenboerger <lechten@wi.uni-muenster.de>]

Dear all,

does someone here produce audio via Text-To-Speech (TTS) from Org
sources?  I plan to do that in the context of emacs-reveal to
generate voice-over for reveal.js presentations, with open questions
[1] concerning my initial, experimental approach.

Currently, I like the default model of Coqui-AI TTS [2] and
Microsoft SpeechT5 [3] best.  Any suggestions for free and open TTS
implementations that produce even better results?  Other models of
Coqui-AI?  The solution should work without GPU support, which seems
to rule out Suno Bark [4].

The above models do not pronounce numbers/digits, and they fail to
pronounce most acronyms.  In a preprocessing step I could replace
those.  I use preprocessing anyways to get rid of Org markup that
might confuse the language models.  Anyone here who did that
already?  Maybe gruut [5] in conjunction with SSML [6] handling?

Any other suggestions?

Best wishes
Jens

[1] https://gitlab.com/oer/emacs-reveal/-/issues/20
[2] https://github.com/coqui-ai/TTS/
[3] https://huggingface.co/microsoft/speecht5_tts
[4] https://github.com/suno-ai/bark
[5] https://github.com/rhasspy/gruut
[6] https://www.w3.org/TR/speech-synthesis11/