3 ms·

So, I've been playing with this a little bit now and here are some comments: - it is very fast and scales quite nicely on CPU with 4 threads (~ twice the speed

So, I've been playing with this a little bit now and here are some comments: - it is very fast and scales quite nicely on CPU with 4 threads (~ twice the speed), but not further (I tried it on a 64 cores box). Not sure why since they seem to be using torch's native threading support. - surprisingly, it is not that much faster when run on a GPU - the quality (as in: what I'm hearing, not a formally measured metric) is good but (YMMV) not as good as turtle. - it breaks with strange error messages if the text you feed it is too long - there is mention of "a model for text repunctuation and recapitalization", which I wonder if it could be used to break a very long text (eg a book) into pieces that can be digested by the tts engine (link: https://habr.com/ru/post/581960/ https://habr.com/ru/post/581960/) - as mentioned in another post, if you want a simple 'convert_this_text_file' CLI utility, you'll have to roll your own or use https://github.com/Grumbel/silero-test/blob/master/silero-test https://github.com/Grumbel/silero-test/blob/master/silero-te... Altogether very nice work, especially the speed.

snakers41 4y ago
Many thanks for a detailed and thoughtful comment. > it is very fast and scales quite nicely on CPU with 4 threads (~ twice the speed), but not further (I tried it on a 64 cores box). Well, practically it does NOT scale even past 6 threads. 64 cores are just overkill, and most likely it will only hurt performance. > Not sure why since they seem to be using torch's native threading support. > surprisingly, it is not that much faster when run on a GPU Probably for the same reason, you can speed up the NN only so much. Realistically it can be made 2-3x faster still. Also currently we abandoned batching, so GPUs are not really required at all. > the quality (as in: what I'm hearing, not a formally measured metric) is good but (YMMV) not as good as turtle. I believe the compute required during training and inference … may differ by 3 or 4 orders of magnitude (!). Also note, that some speakers and languages just sound better due to high quality of source material and the amount of work invested and polish. > it breaks with strange error messages if the text you feed it is too long Well, there should be a warning somewhere, but it works with text no longer than 512-1024 symbols. > there is mention of "a model for text repunctuation and recapitalization", which I wonder if it could be used to break a very long text (eg a book) into pieces that can be digested by the tts engine This model only restores some punctuation marks and capital letters. There are libraries like razdel for this - https://github.com/natasha/razdel https://github.com/natasha/razdel

snakers41 4y ago

Many thanks for a detailed and thoughtful comment. > it is very fast and scales quite nicely on CPU with 4 threads (~ twice the speed), but not further (I tried it on a 64 cores box). Well, practically it does NOT scale even past 6 threads. 64 cores are just overkill, and most likely it will only hurt performance. > Not sure why since they seem to be using torch's native threading support. > surprisingly, it is not that much faster when run on a GPU Probably for the same reason, you can speed up the NN only so much. Realistically it can be made 2-3x faster still. Also currently we abandoned batching, so GPUs are not really required at all. > the quality (as in: what I'm hearing, not a formally measured metric) is good but (YMMV) not as good as turtle. I believe the compute required during training and inference … may differ by 3 or 4 orders of magnitude (!). Also note, that some speakers and languages just sound better due to high quality of source material and the amount of work invested and polish. > it breaks with strange error messages if the text you feed it is too long Well, there should be a warning somewhere, but it works with text no longer than 512-1024 symbols. > there is mention of "a model for text repunctuation and recapitalization", which I wonder if it could be used to break a very long text (eg a book) into pieces that can be digested by the tts engine This model only restores some punctuation marks and capital letters. There are libraries like razdel for this - https://github.com/natasha/razdel https://github.com/natasha/razdel