Voice Question

[surreal pewter]

What if a voice was trained off both the new and old voice sets? I also wonder what if you do ratio's of each to tune it? I imagine it would end up sounding like some eldritch abomination but it would be interesting to test. It may be surprising with good elements of both.

[reef nest]

there's a "questions" tag, use that next time please 🙏

[tawdry herald]

It would make more sense to train a vocoder like HifiGAN or BigVGAN on neuro's current voice then run whatever TTS Vedal plans to use for the expressiveness through it. Whether or not the TTS local will determine the steps in the pipeline such as unnecessary conversions from wav to mel and vice versa.

[hot quiver]

It would be better if vedal adjust the voice bit by bit every neuro stream. That way viewers won't notice a thing, then eventually they can adjust to the new voice without noticing a thing. It will probably take atleast a month of stream

[hidden storm]

unfortunately it is not this simple 😔

[vagrant socket]

can neuro use the new voice for friday

[hidden storm]

nope, sorry

[vagrant socket]

ok

[quick pecan]

IMHO this can definitely do it https://github.com/152334H/tortoise-tts-fast, as suggested above it uses BigVGAN for expressiveness, however for singing this other one works well since you will pre-record it before hand:https://github.com/34j/so-vits-svc-fork

[tawdry herald]

tutel tts

[tawdry herald]

i looked at tortoise and it seems that it only outputs at 22050hz

probably because of bigvgan

[quick pecan]

wait what.. you can hear above 20,000 hz? https://www.youtube.com/watch?v=PAsMlDptjx8

[tawdry herald]

no

sample rate of 22050hz

thats like half of the usual 44.1khz or 48

[quick pecan]

you can change it here: reference_clips = [utils.audio.load_audio(p, 22050) for p in clips_paths]

[tawdry herald]

is it that easy?

pretty sure the input and output size of the model need to change as well

cuz the tensor size changes

[quick pecan]

but its tts, it cant get any better. its a waste

[quick pecan]

yeah you need a model to train it with, thats why for tts its a waste?

[tawdry herald]

Not really

I mean neuro can generate both dictionary and audio

[quick pecan]

yeah i found same voice in azure, you can make it sing with SSML tags

[quick pecan]

but for singing its better to use https://github.com/34j/so-vits-svc-fork has the best vocals cloning. coz you can record the performance anyway,, audience will never know the difference

[maiden marten]

Interesting

[tawdry herald]

not the current opensource SOTA