Ok so, Galaco Talk won’t be included with NEO after 15/2/23, and the end of support for her authorisation and deactivation will be on 31/3/24… this is like Galaco prize all over againUpdate on Galaco Talk situation. Don’t fully understand it
Announcement of “galacoTalk” distribution and support end date - VOCALOID - the modern singing synthesizer -
We would like to inform you of the distribution and supwww.vocaloid.com
Won't lie, it doesn't sound like a person at all, but it's still cool to know this existsJust discovered this vsynth/tts called poino. It looks like it can talk and sing
Going by the poino site, that's because there are no voice actors involved. Similar to Adachi Rei, who also has no voice provider. The site says this is done by editing the formants from envelopes of Fourier transforms (or it uses Fourier transforms to synthesize the sounds? I don't really understand the technical aspect of this, haha) - either way, it'll probably never sound perfectly clear or human, but it's a valiant effort.
There are 2 characters: (by the file names in the repo) Laychie and Layney.
A good article on how filtering with the inverse fourier transform works:
Damn, this is amazing! I don’t understand most of it, but I get the basic differences, and I’m amazed at how much work goes into these things.A SIGNAL PROCESSING MOMENT?! ON MY FORUM? IT'S MORE LIKELY THAN YOU'D THINK!
The inverse fourier transform generates the output signal in real time. Fourier transforms are basically the most common way of interacting with signals in signal processing. On the most basic level, a fourier transform converts a signal from the time domain to the frequency domain. It's most easy to understand by looking at a picture:
Before the fourier transform is performed, we see a constant signal being generated. But it's kind of hard to understand and modify in this form.
After we perform the fourier transform and convert the signal into the frequency domain, we see an impulse at the frequency the signal is at. In this case, it's about ~3 hertz (Hz). Super obvious!
If you have an input signal that is changing frequency (pitch) over time, like a human singing voice, we can see the individual frequencies with the fourier transform!
View attachment 7276
What an artificial voice like poino does is filter a signal at a given pitch with the fourier transform to make it resemble a human voice.
We can tell what frequencies certain sounds make up like this:
View attachment 7277
Article explaining it: Identifying sounds in spectrograms
When the spectrogram is red, there's a higher density of sound there at a certain frequency. So we want to amplify those red parts of our signal to create an "a" sound, and filter out the rest.
The fourier transform is our tool for accomplishing this. After filtering and amplifying different parts of that simple sine wave we started with, we can end up with a sound that sounds more like a human voice after inverse fourier transforming it back to the original time domain, which you then play out of your speakers as sound.
A good article on how filtering with the inverse fourier transform works:
Intro. to Signal Processing:Fourier filter
It can be a lot to get your head around at first!
I think the more poignant difference between Adachi Rei and this software is that Missile created Rei's voice samples by hand by manually shifting the source sin wave around in Audacity and playing with every sound manually, which he then exported. On the other hand, this software is generating a voice completely algorithmically. Rei's voice is going through a lot of different layers of processing by the time you export it from UTAU, while poino generates the voice from scratch in real time. I like Rei's voice for sitting squarely in the middle between completely algorithmically generated sound and being lovingly crafted by a person by hand.
Nerd! Kkkkkkkkk <3A SIGNAL PROCESSING MOMENT?! ON MY FORUM? IT'S MORE LIKELY THAN YOU'D THINK!
The inverse fourier transform generates the output signal in real time. Fourier transforms are basically the most common way of interacting with signals in signal processing. On the most basic level, a fourier transform converts a signal from the time domain to the frequency domain. It's most easy to understand by looking at a picture:
Before the fourier transform is performed, we see a constant signal being generated. But it's kind of hard to understand and modify in this form.
After we perform the fourier transform and convert the signal into the frequency domain, we see an impulse at the frequency the signal is at. In this case, it's about ~3 hertz (Hz). Super obvious!
If you have an input signal that is changing frequency (pitch) over time, like a human singing voice, we can see the individual frequencies with the fourier transform!
View attachment 7276
What an artificial voice like poino does is filter a signal at a given pitch with the fourier transform to make it resemble a human voice.
For example, we can see what frequencies make up the vowel "a" like this:
View attachment 7277
Article explaining it: Identifying sounds in spectrograms
When the spectrogram is red, there's a higher density of sound there at a certain frequency. So we want to amplify those red parts of our signal to create an "a" sound, and filter out the rest.
The fourier transform is our tool for accomplishing this. After filtering and amplifying different parts of that simple sine wave we started with, we can end up with a sound that sounds more like a human voice after inverse fourier transforming it back to the original time domain, which you then play out of your speakers as sound.
A good article on how filtering with the inverse fourier transform works:
Intro. to Signal Processing:Fourier filter
It can be a lot to get your head around at first!
I think the more poignant difference between Adachi Rei and this software is that Missile created Rei's voice samples by hand by manually shifting the source sin wave around in Audacity and playing with every sound manually, which he then exported. On the other hand, this software is generating a voice completely algorithmically. Rei's voice is going through a lot of different layers of processing by the time you export it from UTAU, while poino generates the voice from scratch in real time. I like Rei's voice for sitting squarely in the middle between completely algorithmically generated sound and being lovingly crafted by a person by hand.
