Sound Cards vs. Harpsichords

The sound card people make it seem so easy - just click 'harpsichord' then 'play'. But, no one who loves the sound of real harpsichords is fooled by the result!

Sounds are built by most sound cards in three sections:

  1. an initial sound (the pluck) that is assumed to be brief,
  2. a linear decay of a short sound that can be repeated (looped) without hearing the loop, then
  3. a fixed 'termination' sound (which can only be started from section 2.)

But, no real harpsichord behaves this way!

The 'initial' sound is very long in good harpsichords. They have a hundred small strings between the bridge/nut and pins, which soak up and re-radiate all sounds of the instrument. A good harpsichord takes several seconds to 'settle down' after a note is sounded, longer than the duration of most notes in performance. That's an important part of a quality sound.

Once a steady sound has been reached, all instruments that are activated by an inital action, such as a harpsichord, decay exponentially, not linearly. It's basic physics.

And, the termination sound of harpsichords is mostly related to the sound that exists at the termination - it's not fixed. The sound of the plectrum passing the strings is very different for a recently plucked string than for a still string.

So, to recreate the sound of a real harpsichord on a sound card, you must start by deactivating most of the stuff put there by the computer geeks.

For my sound font, the only way I found that worked well enough to satisfy me was to record a single string from pluck to inaudibility. When the sound of the string became comparable to the noise, I applied a ramp to zero in the sample, then terminated the sample. When you 'play' a note with my font, the sample starts; when the note ends, a linear ramp to zero of the sample is activated. There is no true termination sound, because the sound card wouldn't let me provide one that sounded anything like my instrument. It's true that Flemish instruments were designed to have as little termination sound as possible. It's also true that harpsichordists spend hour upon countless hour practising their key release technique so as to minimize termination sound. But, zero termination sound isn't real, it's just forced by most sound cards.

For my instrument, which is very regular from bass to treble, I found that three samples sufficed to recreate the sound adequately. (Fortunately - my 1985 card didn't have memory for more.) Most harpsichords vary more between bass and treble, and will need more samples. Fortunately, modern cards have enough memory that this is no longer a problem.

How do you record a sample? Begin with a microphone and tape recorder with superb transient accuracy. I ended up using a plain ('cosine directionality') 1/2" condenser microphone. Smaller microphones have a higher noise level, larger ones I tried had audible transient problems. All 'directional' microphones, of whatever type, depend on phasing to achieve directionality - the phase of a harpsichord pluck is discernable right to the limit of human audibility, so they don't work. No 'Dolby' or equivalent system can be active in the recorder - they all modify phase of transients too. Then, find the physical location of the microphone that balances quality sound with sound level. In my case, I ended up with the microphone hung vertically down approximately over middle c of the soundboard bridge, at a height about equal to the width of the soundboard. (Several recording engineers have confirmed this as generally the best location for harpsichords.)

Then, tune your instrument to the pitch at which it sounds its best, record one note at a time with a perfect clean pluck, and wait until each sound dies into total inaudibility before doing the next one. For single string sounds, record notes about a tenth apart from one end of the compass to the other. The exact tuning of each note is not important because you will trim it in the soundfont editor.

Digitize the samples into the sound card, adjust the points at which the card shifts from one sample to the next so that transitions are barely audible, match them to A440, then fine-tune the samples so their pitch matches up perfectly at the transition points. At the end of each sample, create a loop segment of zero sound (otherwise some cards will repeat your sample over and over during long notes). Then, add a termination ramp-to-zero function that sounds most like the damping of your instrument.

Of course, if you listen with the ears of a musician, you will repeat this process many times, until you regretfully admit that you have reached the limit of what can be done with current sound cards and seriously consider throwing the whole shebang out rather than publishing it. Don't! If it is the best you can do, publish it. There are tens of thousands of sound cards out there for every real harpsichord. How will anyone ever know how beautiful our instrument is unless we tell them!

John Sankey
other notes on computing