Comment by nonrandomstring | Hacker News Reader

Comment by nonrandomstring | Hacker News Reader

Parent Full thread

nonrandomstring

Never had the pleasure of a Sennheiser but when working in radio I got my hands on a lot of rack vocoders for doing branding, stings and idents. Funny how the number 9000 comes up a lot, like Roland VP9000 and Eventide H9000. 80 and 90s vintage ones like Korg VC-10 or Elektronik EM-26 had unique sounds, but tbh the modern digital recreations are amazing models. There's not a world of difference between vocoding, autotune, shifting, harmonising etc once you realise how all the fx are now based in FFT, convolution etc - just different variations on processing and control graphs - and so it's fun to create your own vocal effects in things like Max/MSP/PureData. Technically there's a distinction between "effects" and "processing" in terms of how much of the direct (parallel) signal is put through. Chers Believe is a yardstick for "effect", whereas a lot of what I hear with Daft Punk (and Air, Kraftwerk) is quite heavily processed as to disguise the original voice entirely - just letting a bit of top/sibilant through to define the stops and fricatives.

Analog vocoders are only nominally like analog FFTs.

The shape of the filters, the smoothing between the filters and the synthesis section, and (on some models) the patchability all create a very different result.

The reason the best analog vocoders are so expensive is because the filter for each band is much more complex than a plain old bandpass filter, with a much higher component count. Typically there's a flatter passband and a steeper slope than you'd expect.

You can do digital convolution with thousands of bins and it sounds nothing like analog vocoding. It's much cleaner, doesn't have those lovely harmonically spaced filter resonances, and creates sounds that can feel more acoustic than electronic.

Did you listen to the example audio in the video? Soft synths and digital emulation can be absolutely amazing these days, but the VSM201 and Ultimate VoIS are in their own league. It’d be pretty easy to pick them out from a blind test with other vocoders.

Oh, it also might be of interest that the IVL algorithm isn’t FFT-based. I think their harmonizers sound better than the rest, so maybe FFT isn’t the best way to go.

I thought it was phase synchronous overlap add, but I just checked and now I'm not so sure.

Has anyone got more details?

That gives me something to research! I’ve only scratched the surface of IVL’s algorithm, but intend to look into it further.

Yes exactly, I was really excited when I found out that you do not need a FFT to do speech processing.

If you look at the code of (phone/voice) codecs GSM/Speex/Opus you can see that you can estimate the spectral envelope (or the configuration of a physical tube model for the vocal tract) in time domain with linear prediction coefficients (LPC).

And it is simple, e.g. the often used Levinson-Durbin algorithm is just 22 lines of C code. It is an interesting exercise to build your own vocoder from scratch that fits in a single screen page.

Many of the code snippets I have seen (which likely have already processed your voice) are just translations of the Fortran code of the book "Linear Prediction of Speech" by Markel and Gray (1976).

Ah yes, ladder or lattice filters. If you don't mind old fashioned mailing lists there's still a few of hanging around in MUSIC-DSP@LISTS.COLUMBIA.EDU where code gets shared.

Reply on news.ycombinator.com