DSP Robotics Support

Ahh so I just miss the place. SInce coefficient calculation is hoped it is better to implement this fix in that code. Here it is (not optimized becasue it is hoped).

Nice!
That block of min/max primitives can be deleted too - the min and max is already there in the code.

Thanks, I do the checkup on how it behaves (in the whole schematic) tonight. Who would like to mess around the optimization a little bit? MyCo?

There is one unanswered question I have here. I asked it few times on SM forum, but nobody responded. If I feed the format filter with tonal input (like saw) - then the voice pitch is easily regulated.

But what about the whisper? Noise has no particular fundamental pitch. So feeding this filter with noise - there is only one whispering "pitch". Can be compared to resonant frequencies of the vocal tract. But humans (male/female/kids, etc) different features there, so the whispers can differ from each other.

So I was wondering - is there a quick way to manipulate pitch of the whisper, without affecting proper (human-like) formant structure? Would it be by adding some sort of filter to input noise, or different tables/coeffs for vowels inside format filter?

I'd guess that shifting the formants your best bet - like transposing an instrument, keep the relative distances between them the same, but move them all up or down. That's probably as close as you'll get to a "biological" model - and it would work on the pitched sounds too, to model the different size of people's vocal tract.

I had a synth plugin that worked that way - there were no "normal" oscillators, just white noise sources and resonant band-pass filters that tracked the notes played. Can't remember now what it was called; I'll have to see if I still have it, it made some really beautiful pad sounds quite unlike any oscillator based synth.

EDIT) Aaah, here it is... Chimera

Okay, thanks for hint. My new little fellow will be also pretty playful when I finish it.

Martin - quick question, because that vowel filter was your design.

Do you have a recipe for continuous vowel-like sounds, i.e. L, M/N, R?

tester wrote:a recipe for continuous vowel-like sounds, i.e. L, M/N, R?

You could try google for the respective formant frequencies and intensities. With that data it would be possible to augment the set of filter coefficient tables that are already there for a,e,i,o,u.
Another option would be to create snapshots of the synthesis filter coefficients in my LPC schematic here when feeding these sounds into the voice input. This way you could actually create a personalized catalog of your own vowels and other voiced phonemes. That's something I wanted to do myself one day, just need to get a good mike.

But these semi-vowels will appy as tables to the same schematic you made? Isn't this something applied after? These sounds are made a little bit differently, that's why I ask.

This LPC seems to be an introduction to personalized text-to-speech synthesis.

Another question I asked before, but can't remember if anyone answered. When feeding the formant filter with noise (whispering vowels) - what sould I do to achieve different pitch at the same vowel? What sort of filter where and how?

I see if I can find anything that I uderstand.
Thanks!

My advice would be to listen, you can descern speech.. So if you can tune two or three filters to form the vowel you want it cant really be wrong.

For changing the pitch you can raise or lower the filters a little, as long as the underlying ratios are more or less the same. And it still sounds like the letter you want.

I have on a synth what trog said, a compressing and expanding ratio. It does have a cool yeaaah aieeee sound.

For speech processing - certain things must be specifically in tune. Otherwise speech will be not higher or lower pitched, but higher or lower resampled (which isn't sounding well at octave difference). I don't know which parameters to tune in Martin's work.

Plus - I assume, there are already existing models, that offer tables with coeffs/parameters. I'm not an expert in that field, but I'm interested in specific results. Martin knows better what to do with the theoretical stuff and how to translate it into something I can understand.