Thank you so much for making the changes! It does sound much more fluid now. However, I've looked at the code on line 190 that sets the silence duration and experimented a bit with various numbers. In the runs I did, .7 was the most intelligible. With .6 I sometimes lose track of what its reading. With .8, it sometimes introduces longer pauses in the middle of a sentence, typically at a comma or near an "and", shifting also a bit the intonation as if a new sentence were started.

Another observation, which is independent of the pause length parameter - whenever there is a sentence longer than 240 notepad columns (incl. white spaces), one of two things happens:
a) there is an abrupt end to the sentence around column nr. 240 and the rest of the text is read as if a new sentence is beginning, thereby producing a non-natural intonation
b) the voice starts to mumble something or repeats the last two-three words, is silenced and then continues reading as though from the middle of the sentence (without changing intonation)
It seems as though coqui is unable to read sentences longer than 240 and force cuts them at this point. To prevent this from braking the flow of the text, maybe it would be possible to split the sentence at the first comma or before the first "and" that precedes column for ex. 230? (I'd suggest not waiting until it hits its buffer of approx. 240, but splitting before 230 or 220 to give a bit of leeway to the model).

And a final, third, observation - I notice that it frequently doesn't read the final letters (usually "s" and "t") of the sentence. Is this coqui related, or do you have code in epub2tts that somehow silences the voice at the end of the sentence?

The key message though is: great job on this update! It very much improves the quality and legibility/understandability of the text being read.

Pause between sentences #208

Description

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions