2.1 The Instrument Database

SPORCH takes as its input parameters a source sound file and a specification for a group of instruments. Any instruments may be included, as long as sample recordings of pitches at various dynamic levels over their entire ranges are available.

Before the program can be run, a database containing spectral analyses of all instruments must be built. The user must first supply all of the sound samples and create a “database specification file” that supplies acoustic data and tells the program how it should analyze the instrument samples. SPORCH automatically interpolates missing data within a reasonable distance if all pitches or dynamics are not available. An instrument may have several techniques associated with it (for example, a string instrument can have separate data for both arco and pizz. playing techniques). Percussion and other non-pitched instruments may also be included—the default analysis procedure (extracting peaks over the entire frequency range of the spectrum) handles both pitched and non-pitched sound sources.

The compiled database contains the following information for each instrument and technique:

• The pitch range (if applicable).
• A reduced or “practical” pitch range (if applicable), specifying the pitches that are generally easier to play. This is used by the orchestration algorithm to produce parts that tend to stay within a comfortable range rather than leaping to extreme registers. The parts are then more playable than they would be otherwise.
• A collection of the most prominent peaks (i.e. spectral maxima) taken from a spectral analysis of each pitch at various dynamic levels. If information for some pitches or dynamic levels is missing, peaks from neighboring ones are transposed and/or amplified to interpolate missing data.

The bulk of SPORCH's code is dedicated to building this database and interpolating the large amount of data. The user controls several analysis parameters so that the data can come from either the steady state portions of the tones or the attack portions if this is more appropriate (as in staccato articulations, for example).

To get from digital sound samples to sets of peaks, sporchdb, SPORCH's database compiling program, uses the following procedure:

• A series of FFT analyses are taken over a specified time range.
• The highest amplitude components that are present in each of the FFT frames are extracted so that noise and leakage in the FFT analysis are eliminated. In addition, peaks are extracted until the entire audible frequency range from 20 Hz to 20,000 Hz is covered to within the error margin described below (see The Search). This is a simple model that includes both the highest peaks with accurate frequencies and lower ones that include the noise that occurs in the sound. From all of the peak sets, the ones that “deviate” the most from the other sets are discarded. Deviation is determined by finding distances between matching peaks and calculating the root mean squared error between them.
• The remaining sets of peaks are averaged together to form the final representative set. This is done by averaging the amplitudes and frequencies of matching peaks.

Several post-analysis modifications may be done. For example, peaks may be tuned to certain fundamental frequency (necessary for some of the free sample databases that are available) and/or amplified. The same entire procedure is also used when the program is run and a sound source is analyzed.