Suggestion: Preserve commas in Kokoro's preprocessing.

[LostRuins]

I've been playing around more with Kokoro ever since I've integrated it into KoboldCpp, and I've noticed that you strip out commas , for dashes -- in order to create a pause.

TTS.cpp/src/models/kokoro/model.cpp

Line 1416 in 724d97f

normalized = replace_any(prompt, ",;:", "--");

Although it might seem similar on surface, the comma contains useful information that affects the characteristics of the generated audio.

Consider this sample phrase:
stealing from unsuspecting travelers, picking pockets, and conning the locals

There should be two pauses in a proper narration. This is from the hexgrad space

good.mp4

Now compare this with the TTS.cpp output, which since it strips all commas and replaces them effectively with spaces, generating this:

bad.mp4

Adding commas back solves the issue. I'm not sure if I missed something, so feel free to correct me. I do not use espeak so I don't know if there are negative repercussions there, but without espeak this results in an overall better narration for me personally.

[LostRuins]

Also, this is another very dirty hack I wanna share (feel free to ignore haha) but I found that adding commas after every single sentence's punctuation actually improves coherence for non-espeak. I don't know if you'd find it useful, but observe my no_espeak generation for Kokoro.

This is an unedited no-espeak generation.

compare.mp4

Source text:

Feel free to check out my non-espeak hacks and adopt any of them you like! I know they are pretty crazy, but I just kept messing around until I found something that suited my use case.

LostRuins/koboldcpp@ff1d179

Finished IPA for your reference

luːwˈiːzᵻz ˈIz wˈɛld wɪð mˈYsʧəɹ æz ɐn ˈænəmᵊl kɹˈI ʃˈʊk hˌɜɹ hˈaʊs, ænd ʃi fˈWnd həɹsˈɛlf ʃˈɪvəɹɪŋ æz ɪt dˈId əwˈA., ʃi ʃˈɪvəɹd ˈɪn ɐ vˈɛɹi dˈɪfəɹənt wˈA wˌɛn 
ʃi wˈIpt æt wˈʌn ˈI ænd hˌɜɹ fˈɪŋɡəɹtˌɪps kˈAm əwˈA kɹˈɪmzᵊn wɪð blˈʌd., ɪts ðə həlˌusənˈAʃənz, ʃi tˈOld həɹsˈɛlf., ʃi wˈɪnst ænd hˈɛld ɐ hˈænd tu hˌɜɹ lˈoʊɚ bˈæk æz ʃi ɹˈoʊz tu ɐ stˈændɪŋ pəzˈɪʃən, blˈɪŋkɪŋ ðə blˈʌd ˈWt ʌv hˌɜɹ ˈIz., hˈæz tu bˈi., əɡˈɛn, ðə kɹˈI ˈɛkOd θɹˈu ðə tˈWn ʌv kˈɛnə, nˌɑt ɐ dˈɔɡ, nˌɑt ɐ wˈʊlf, nˈɔɹ kˈIOt., ɪt 
ˈɛkOd æz ðˌO ɪt kˈAm fɹʌm fˈɑɹ əwˈA, bˈWnsɪŋ ˈɔf ʌv ðə nˌɪɹbˈI mˈWntᵊnz, bˌʌt ɪt hˌæd ɐ vˈɑljˌum bˈɛTəɹ sˈuTɪŋ sˈʌmθˌɪŋ ʤˈʌst aʊtsˈaɪd ʌv luːwˈiːzᵻz hˈOm., ðə dˈɑktəɹz 
sˈɛd ˈaɪ mˌIt sˈi ɔɹ hˈɪɹ θˈɪŋz, ʃi ɹəmˈIndᵻd həɹsˈɛlf., wˌɛn ðə bˈɑːdɪz ˈɪn ɹˈʌf ɪnˈʌf ʃˈAp, ðə bɹˈAn stˈɑɹts tu ɡˌO tˈu., ɪt wʌz ɐ stɹˈAnʤ mˈɪks ʌv əmˈOʃənz, fˈilɪŋ hˌɜɹ hˈɑɹt ɹˈAs æt ðə sˈAm tˈIm ʃi fˈɛlt æt pˈis., ʃi hˌæd bˌɪn bɹˈAsɪŋ həɹsˈɛlf fɔɹ ðɪs sˈɪns mˈʌnθs əɡˈO.

P.s. yeah I am avoiding espeak so I can run this without any added dependencies, and not worry about other licenses.