TGSpeechBox v3.0-beta2 is here

[tamasg@mindly.social]

TGSpeechBox v3.0-beta2 is here!
Big changes: SAPI is now a single DLL! No more juggling speechPlayer.dll, nvspFrontend.dll, and libespeak-ng.dll. Runtime code cut nearly in half.
Per-voice settings on mobile! each voice remembers its own tuning when you switch.
Diphthong quality leap: bandwidths now interpolate alongside frequencies during glides, fixing the "shaky" vowel quality reported by users. Onset settle time lets resonators establish before the glide begins. Adaptive hold scales with formant distance so narrow diphthongs don't stall and wide ones don't smear.
Fixes: limiter no longer pumps at pitch rate (goodbye "same page" grittiness), stops no longer sound like affricates at low volume (/t/→/tʃ/ and /p/→/f/ gone), quote-aware clause splitting fixes intonation through dialogue, and iOS clicking from rapid VoiceOver swiping is resolved.
Also: GOAT monophthongization, velar word-final boost, sample rate picker on iOS, pause mode on Android, and a bunch more.
https://github.com/tgeczy/TGSpeechBox/releases/download/v-300b2/tgSpeechBox-v300b2.nvda-addon
https://github.com/tgeczy/TGSpeechBox/releases/download/v-300b2/tgsbPhonemeEditorWin32-v300b2.zip
https://github.com/tgeczy/TGSpeechBox/releases/download/v-300b2/TGSpeechSapiSetup-v300b2.exe
https://github.com/tgeczy/TGSpeechBox/releases/download/v-300b2/TGSpeechBox-v300b2.apk
https://github.com/tgeczy/TGSpeechBox/releases/download/v-300b2/tgspeechbox-linux-x86_64-v-300b2.tar.gz
https://github.com/tgeczy/TGSpeechBox/releases/download/v-300b2/tgspeechbox-linux-aarch64-v-300b2.tar.gz

[chrisduffley@mastodon.chrisduffley.com]

@Tamasg OK, is this just me or is it sounding more UK/Scotish under certain phonemes? Five sounds more scotish (somewhat), at least that I've noticed the most. I don't know if it really sounds authentically UK now though lol.

[tamasg@mindly.social]

@ChrisDuffley really good one to flag. This is with En-GB? Perhaps some of the tuning to get "eight" to sound better messed with that, hmm. worth looking over the changelog if so

[chrisduffley@mastodon.chrisduffley.com]

@Tamasg Yep indeed. Also the iOS version doesn't use the newer phoneme table NVDA is using, so R's are quite stuck out for instance, in Sierra and Zero for sure. Also, speakking of which, the o in 0 is quite, hmm, I can't even explain it. But it's a bit lighter than it used to be. For some reason I always compare to regular NVSpeechPlayer and ESpeak as they have the best British phoneme table out there, in my opinion.

[chrisduffley@mastodon.chrisduffley.com]

@Tamasg LOL, another one. The "OU" in thousand is quite odd, the end of it is kind of long. Just find it really interesting lol.

[kaveinthran@disabled.social]

@ChrisDuffley @Tamasg While you are using the ngb table, how words like "sharp" "wrap" sounds to you?

[chrisduffley@mastodon.chrisduffley.com]

@Tamasg a few other things. Australia which I switchd to for the first time. The o is so British really. Is the "u" in public suposed to be so, like, I can't describe it? It should have more of a mouth-opening ish uh in it, shouldn't it? Yeah, you're talking to a stickler for accents but can't even properly fix a phoneme

[chrisduffley@mastodon.chrisduffley.com]

@kaveinthran @Tamasg a slightly more opened up o in shop, for how sharp sounds here, and also wrap is a bit more closed I guess, it's really hard to describe stuff like this lol.

[tamasg@mindly.social]

@ChrisDuffley Gosh Australian is going to need a lot of work. I know there's Hilenbrand phoneme tables for US English, but I wonder if anything like that exists for Australian, where they gathered it from actual speakers. Something like that would really help it, right now it's more like a bad mix of US and UK, which was just about as much as I could pull out of the research, but not all the patterns where it sounds which. Like, flaps are more US, but vowel centralization definitely different and can be US on some. So yeah, it's honestly like, one of the trickiest languages, but I hope I can improve it. The other UK issues might be related to diphthong collapse though, I'll try to disable it and see because it does do some shifting of the bandwidths, and if it's shifting the onset of that sound it could absolutely become more Scottish. Same with something like Thousand where you glide the Ou-w sound together a bit.

[kaveinthran@disabled.social]

@ChrisDuffley @Tamasg At least for me, in beta 1, I couldn't here the "ep" sound in those words. US english works good though. When you compared with espeak, you can hear the "ep" sound. A fun way to do this is to allow both espeak and speech box read the first 10 sentences or first page of the first book of Harry Potter.

[kaveinthran@disabled.social]

@ChrisDuffley @Tamasg In many words with the ngb, the p sound at an end of word is considerably silent.

[kaveinthran@disabled.social]

@ChrisDuffley @Tamasg you also can compare with the classic NV speech player for ENGB, it's the gold standard. I wonder, could we use the NVSpeech Player phonemes for ENGB as it should be already stable?

[tamasg@mindly.social]

@kaveinthran @ChrisDuffley not really. Other languages have altered that by now slightly, reverting them is a bad move. Also there's passes like coarticulation that make a difference on the final sound, even if we did so. It's a lot more complex machinary interacting together.