The plan is to work through the Dīgha Nikāya first before moving on to other collections. I hope it’s useful for anyone who enjoys listening to the suttas.
=======================
Edit:
I did use TTS synthesis to generate the audio since I can use different voices to alternate. That helped me better engaged while first going through the long texts.
I’m sorry about posting something banned on this site.
To be clear, I did not use the content to “feed AI models” or whatever.
Hm… OK. I’ll delete the post.
I feel a little hurt since I didn’t use AI mindlessly. I reviewed the audio unit by unit and the result does help me better engaged in listening the sutta. That’s what made me thought maybe someone else might find it useful.
Plus, TTS is not “feeding the AI”. Anyways, I won’t argue.
I agree that it’s not as specific as could be, partly because the definition of “AI” can be quite broad.
SC Voice, as I understand, is partly exempted as it uses “AI” for accessibility purposes[1].
I may be mistaken, but I believe the variant of Algolia that the instant search uses is DocSearch, which predates the advent of LLMs.
If you want I’d be happy to read with you like post a reading of pali, I’ll post a reading and then see how far can go in something - maybe as go on learn improve pali pronounciation, etc
EDIT: just to be clear I don’t know pali (just tiniest smidgen not even right to say “A1” if you know what a conventional language knowledge is) just as practice (because it’s awesome) etc EDIT2: At the moment I’m most keen on KN and jump to Itivutikka and kinda cool example as where lay person of incredible parami has been asked by the sangha to recall the dhamma or sati sutta in DN (dn22) the main one for me rn dn33 “sparks joY” as they say in the past though it was mostly reading about SN with what struck a chord
Thank you all. This post became a chance for practice for me.
I appreciate all the feedback, either positive or negative or neutral.
I watched my mind’s reaction to critical comments and comforting comments. When “my work” is devalued, I felt I was devalued. There is an attachment to things we put effort into. The hurting feel revealed that.
When there is comforting comment, there is a sense of “At least, I’m not rejected by this community as a person”. Again, sense of me and attachment to this community. The wanting of being accepted and liked by this community.
At end of the day, I myself is responsible for letting go of all these graspings.
SC-Voice also uses AI to estimate audio lengths so that users have some idea of the length of the audio. In that sense, it is “generative” in the way that “taking a multivariate average” is generative. The SC-Voice audio neural net generates an approximation of the real audio length. Statistics is the science of approximations, which are all computed averages that use special formulas that highlight various relationships in the data. I saw no point in spending lots of money recording the actual audio length of each text given that the translations are constantly being updated. So yes, the audio length is indeed generated by AI. And users typically want to know how long the audio is before they click play to generate the audio using TTS, which is also generative AI.
I wrote the code for a neural net from scratch to understand how gradient descent worked. It proved to me that all a neural net does is transform the data into a multidimensional space where training is basically just finding a straight line through points in that multi-dimensional space.
I tend to use libraries sparingly so that I can get a deeper understanding. I found that simply plugging in a library never gave me a sense of the tradeoffs or implications in the use of a library. A similar issue applies to the use of AI. If we use AI blindly, we engage in a shallow relationship with the world. Expecting AI to give us “truth” is like expecting a straight line to be reality. The error is in the grasping expectation, not the line.
The category of “AI” here sounds like a classifier to me. I.e. old school ML (Machine Learning), not the “Gen AI” of current times.
Using deep learning (neural nets) seems more complex than necessary for this kind of classification/estimation to me, but of course, accuracy would be the main barometer to measure against.
I hadn’t read through enough of the codebase to understand how the audio was being stored. Audio files usually have length as part of the metadata, but I’m guessing you’re chunking it differently or AWS Polly doesn’t give back length in the specific API you’re using?
Also, sorry, this is getting pretty off-topic (perhaps should be split off), I mainly wanted to make the clarification around classifiers
As far as I understand it—but my understanding is certainly incomplete—the problem here is that audio is generated and stored by segment, not by sutta. So it’s not so easy to know how long it would take to play a specific sutta.
But I am indeed impressed that in this short time since you got interested in the projects around SuttaCentral you managed to get quite a good understanding of things!
Ah, that makes sense, especially given SC’s segmented interface.
I’d think you could get an estimate for a minimum by adding together the lengths of all segments within the sutta (for a single language). I’m not sure how SC-Voice handles the pauses in between segments, but that would have to be added on and may be more variable, I’d think.
I’m guessing there’s something else that I’m missing that makes it more complicated than that?
Thank you for your kind words Learning quickly tends to be one of my talents, although I’m much faster in breadth than depth. Very deep concepts still take me significant time – otherwise the more “hard to understand” pieces of Dhamma would have clicked by now
I should say I’m impressed by your knowledge of the technical nuances without having been a software engineer. That and your very detailed feedback would probably make a pretty good career for a product manager if you weren’t in robes
I’ve been closely involved in developing SC-Voice together with @karl_lew, discussing together in which direction features should go, testing and the like. We had many discussions on what is and what isn’t technically possible, and with how much effort, etc. Most of my knowledge of technical things I’ve learned from Karl.
However I totally lack any depth in technical knowledge.
Learning quickly tends to be one of my talents, although I’m much faster in breadth than depth. Very deep concepts still take me significant time – otherwise the more 
