Descriptors

These settings shape the audio measurements used for onset detection, pitch estimation, and AI model input. Keep them consistent with model training. Setting a descriptor parameter does not itself request that descriptor; the host API, score events, and model inputs determine which optional analyses are needed.

Audio analysis

Keyword	Default	Accepted value	Effect
`SR`	`48000` Hz in parsed configuration	Positive integer	Analysis sample rate. Match the host audio rate and model training rate.
`FFTSIZE`	`2048` samples	Positive power of two	Analysis window size. Larger windows cover more audio and provide finer frequency bins.
`HOPSIZE`	`512` samples	Positive power of two	Spacing between analysis updates. Smaller hops request more frequent updates.

SR does not resample the audio. Score loading reports a mismatch with the engine's existing rate. Set it explicitly when using a rate other than 48000 Hz.

Integer settings are converted from numeric values before validation. Use whole numbers. The parser validates each option separately; it does not check every relationship between settings.

Onset and articulation

Keyword	Default	Accepted value	Effect
`ONSETFUNCTION`	`mkl`	One of the lowercase names below	Selects the onset detection function.
`MEDSPAN`	`50`	Positive integer	Median span used by the onset detector.

Value	Implementation selected by the score parser
`pow`	Power (`ODS_ODF_POWER`)
`pd`	Phase deviation (`ODS_ODF_PHASE`)
`wpd`	Weighted phase deviation (`ODS_ODF_WPHASE`)
`sf`, `hfc`	Both select magnitude sum (`ODS_ODF_MAGSUM`)
`cd`	Complex domain (`ODS_ODF_COMPLEX`)
`rcd`	Rectified complex domain (`ODS_ODF_RCOMPLEX`)
`mkl`	Modified Kullback–Leibler (`ODS_ODF_MKL`)

In the current parser, sf and hfc are equivalent selections. Switching between them does not select different algorithms.

Test onset changes with the exact articulation that fails. Compare attack detection before changing FFTSIZE or HOPSIZE, which also affect other analyses. See Onset Descriptor for the output.

MFCC and logmel

Keyword	Default	Accepted value	Effect
`MFCCMELS`	`40`	Positive integer	Number of mel bands used by MFCC and logmel.
`MFCCCOUNT`	`13`	Positive integer	Number of MFCC coefficients.

These settings change model input dimensions. Match the trained feature extraction setup; changing them is not a general remedy for poor technique recognition.

YIN pitch estimation

Keyword	Default	Accepted value	Effect
`YINTHRESHOLD`	`0.15`	`0` to `1`, inclusive	Threshold for the first candidate minimum in YIN's normalized difference function.
`YINMINFREQUENCY`	`50` Hz	Positive number	Lower frequency limit for estimated pitch.
`YINMAXFREQUENCY`	`2000` Hz	Positive number	Upper frequency limit for estimated pitch.

Choose a minimum below the maximum and a range supported by the sample rate and window size. The parser checks positivity separately, without enforcing the relationship between the limits. YIN supplies pitch descriptors; harmonic template matching has its own Pitch Following settings.

Spectral rolloff and chroma

Keyword	Default	Accepted value	Effect
`SPECTRALROLLOFFCUTOFF`	`0.85`	`0` to `1`, inclusive	Fraction of spectral power used to locate the rolloff frequency.
`CHROMASIZE`	`12`	Positive integer	Number of chroma bins.
`CHROMACENTEROCTAVE`	`5`	Any number	Center octave of the chroma weighting.
`CHROMAOCTAVEWIDTH`	`2`	Positive number	Width of the chroma octave weighting.

Zero-crossing rate

Keyword	Default	Accepted value	Effect
`ZCRCENTER`	`true`	Boolean	Enables edge-value padding for centered ZCR analysis.
`ZCRPAD`	`false`	Boolean	Counts an initial crossing when enabled.
`ZCRZEROPOS`	`true`	Boolean	Uses sign-bit comparisons; when disabled, zero is treated as a separate sign value.
`ZCRTHRESHOLD`	`0.0000000001`	Non-negative number	Samples with absolute magnitude at or below this value are treated as zero.

For these three boolean options, the parser accepts true, on, yes, or 1, and false, off, no, or 0. Word values are case-insensitive. ZCRPAD controls the initial crossing count; ZCRCENTER controls edge padding.

Example

BPM 72
SR 48000
FFTSIZE 2048
HOPSIZE 512
ONSETFUNCTION mkl
MEDSPAN 50
MFCCMELS 40
MFCCCOUNT 13
ZCRCENTER ON
ZCRPAD OFF
ZCRZEROPOS ON
ZCRTHRESHOLD 0.0000000001

NOTE C4 1

This illustrates configuration syntax using defaults. For descriptor meanings and output values, see Audio Analysis. For input selection and metadata precedence, see AI Model Recognition.