@Tamasg Just wanted to congratulate you on version 2.99 of TGSpeechbox! The impulse pitch mode is almost, almost eloquence now. The only real issue is the "o" sounds in words like social and mode being a bit too long haha. But I can now go as fast as eloquence for predictable work. I've got to slow it down quite a bit for reading, but I'm not sure how much of that is just that I haven't fully adapted to the new voice, yet.
@fastfinge oh this is good to know! Yeah, the impulse rewrite was maybe 5 hours of straight work, lots of ear tuning. Now it's Klatt mode that sounds too flat, but there's languages that might benefit from that kind of prosody too. I just fixed a bunch of English tuning (not live yet, but will be soon) on words like card, shard, party, ETC. Way too fronted, now it's more central. So tuning is a slow drip-feed of boring ear-by-ear listening, plucking a phoneme and making it sound better based on what people describe. Ripping out Espeak's phonemizer only was surprisingly easy, so I'm not having to borrow the entire car just to get the GPS out of it, making it have really good responsiveness there, so I was quite happy to discover that once going MIT. Doesn't make it the best IPA phonemizer, but for sure the most diverse.
@fastfinge ah still Espeak, but not the entire lib and voice processing engine. It now statically compiles straight into the libs on mobile, so all the phonemizer stufff was like 17 C files that get linked into the project. Definitely would have been a big no-no with the GPLV2 V3 mismatch earlier.
