Comment by blincoln | Hacker News Reader

Comment by blincoln | Hacker News Reader

Are there any detailed descriptions available of how the music and voice were synthesized?

Based on the recording, the information I could find, and imagining how I'd try to do the same thing using the technology of the era, I assume the melody is based on single-cycle samples of a piano and Max Matthews playing the violin. The vocals sound like formant synthesis like the Votrax SC-01 or TI LPC series, although of course those chips didn't exist until 15+ years after the work at IBM. But I'm very curious about the details. Did the team develop a general-purpose sequencer for the melody and/or speech, or were all of the notes, slides, etc. hardcoded? Did the computer actually output all 3+ parts together, or were they separate elements mixed after the fact? I assume the output was not realtime, but it would be a neat surprise if they achieved that in the 60s. Was it all handled digitally in the computer, or was the computer controlling some add-on hardware, maybe with analogue filters? Etc.

The singing synthesizer used a surprisingly sophisticated physical model of the human voice [1].

The music was mostly likely created using some variant of MUSIC-N [2], the first computer music language. The syntax and design of Csound[3] was based off of MUSIC-N, and I believe the older Csound opcodes are either ported or based off those found.

Apparently the sources for MUSIC-V (the last major iteration of the MUSIC language) can be found on github [4], though I haven't tried to run it yet.

1: https://ccrma.stanford.edu/~jos/pasp/Singing_Kelly_Lochbaum_...

2: https://en.wikipedia.org/wiki/MUSIC-N

3: https://en.wikipedia.org/wiki/Csound

4: https://github.com/vlazzarini/MUSICV

Thank you! Seems like that project was incredibly far ahead of its time.

The physical-modelling aspect is super interesting. Does that mean that the similarity in sound to formant-based speech synthesis is because they're both using a sawtooth wave, noise, or other relatively simple sound as the raw input? I always imagined that a physical-modelling speech synthesizer fed by a sawtooth wave would sound more like a vocoder than Votrax or TI LPC output does, but I guess not.

> Does that mean that the similarity in sound to formant-based speech synthesis is because they're both using a sawtooth wave, noise, or other relatively simple sound as the raw input?

Essentially, yes. Both are known as "source-filter" models. A sawtooth, narrow pulse, or impulse wave is a good approximation glottal excitation for the source signal, though many articulatory speech models use a more specialized source model that's analytically derived from real waveforms produce by the glottis. The Lilencrantz-Fant Derivative Glottal Waveform model is the most common, but a few others exist.

In formant synthesis, the formant frequencies are known ahead of time and are explicitly added to the spectrum using some kind of peak filter. With waveguides, those formants are implicitly created based on the shape of the vocal tract (the vocal tract here is approximated as a series of cylindrical tubes with varying diameters).

Human speech production/perception works by articulation changing the shape, hence resonant frequencies (formants), of the vocal tract, and our ear/auditory cortex then picking up these changing formants. We're especially attuned to changes in the formants since those correspond to changes in articulation. The specific resonant frequency values of the formants vary from individual to individual and aren't so important.

Similarly the sound source (aka voice) for human speech can vary a lot from individual to individual, so serves more to communicate age/sex, emotion, identity, etc, not actual speech content (formant changes).

The reason articulatory synthesis (whether based on a physical model of the vocal tract, or a software simulation of one) and formant synthesis sound so similar is because both are designed to emphasize the formants (resonant frequencies) in a somewhat overly-precise way, and neither typically do a good job of accurately modelling the voice source, and other factors that would make it sound more natural. The ultimate form of formant synthesis just uses sine waves (not a source + filter model) to model the changing formant frequencies, and is still quite intelligible.

The "Daisy" song somehow became a staple for computer speech, and can be heard here in the 1984 DECtalk formant-synthesizer version. You can still pick up DECtalks on eBay - an impressive large VCR-sized box with a 3" 68000 processor inside.

https://en.wikipedia.org/wiki/Daisy_Bell

I guess they built upon the Voder [1] (Homer Dudley, Bell Labs 1939 as well. But that was played manually, amazing ‘instrument’?

1: https://youtu.be/5hyI_dM5cGo

Sort of. Both use articulatory synthesis, which attempts to model speech by breaking it up into components and using some coordinated multi-dimensional continuous control to perform phonemes (the articulation aspect). The voder uses analog electronics, while Daisy does it digitally (and without a human performer).

The underlying signal processing used for both is different, but both use a source-filter mechanism.

The synthetic voder output sounds more or less exactly like the output of a vocoder where the input is a human voice and the carrier is a sawtooth. Not surprising, given that the voder was made by the same people.

But I'm still unsure why those two things sound so similar to each other, and formant/LPC chips sound so similar to each other, but the two groups of things sound so dissimilar (at least, IMO).

I have a background in electronic music, so I'm pretty familiar with additive, subtractive, and other types of synthesis.

I'm especially surprised about the physical modelling sounding more like a formant chip, because a guitar "talk box" gives a sound exactly like a vocoder, and that should be almost the same thing, just with a real human mouth instead of a model.

The vo(co)der uses banks of fixed filters to apply the broad shape of a spectrum to an input signal. It's basically an automated graphic EQ. The level of each fixed band in the modulator is copied to the equivalent band in the carrier.

The bandpass filters have a steeper cutoff than usual and are flatter at the top of the passband than usual. And the centre frequencies aren't linearly spaced. But otherwise - it's just a fancy graphic EQ.

The formant approach uses dynamic filters. It's more like an automated parametric EQ. Each formant is modelled with a variable BPF with its own time-varying level, frequency, and possibly Q. You apply that to a simply buzzy waveform and get speech-like sounds out. If you vary the pitch of the buzz you can make the output "sing."

LPC uses a similar model but it applies data compression to estimate future changes for each formant band. So instead of having to control all the parameters at or near audio rate, you can drop the control rate right down and still get something that can be understood.

There are more modern systems. FOF and FOG use granular synthesis to create formant sounds directly. Controlling the frequency and envelope of the grains is equivalent to filtering a raw sound, but is more efficient.

FOF and FOG evolved into PSOLA which is basically real-time granulated formant synthesis and pitch shifting.

Many of the simpler vocal tract physical models are very similar to the cascaded allpass filter topologies found in LPC speech synthesizers.

In general, tract physical models have never sounded all that realistic. The one big thing they have going for them is control. Compared to other speech synthesis techniques, they can be quite malleable. Pink Trombone [1] uses a physical model under the hood. While it's not realistic sounding, the interface is quite compelling.

1: https://dood.al/pinktrombone/

Reply on news.ycombinator.com