SpectraScore Orchestration Matcher — User Guide

Analyze the measured spectrum of a selected Sound and derive a playable orchestration: choose an instrument combination, assign pitches from real spectral components, preserve microtonal detail, and export the result as MusicXML and a Praat Table.

Author: Shai Cohen Affiliation: Department of Music, Bar-Ilan University, Israel License: MIT License Repo: Praat AudioTools

Contents:

What this does Sound → orchestration Quick start Analysis engine Measured spectral peaks Instrument model Ensemble search Voicing strategies Controls Outputs Visualization Reading the results Notes & limitations

What this does

SpectraScore Orchestration Matcher treats a recorded or synthesized Sound as an orchestration target. It measures the Sound's spectral shape, detects a reference root, extracts significant spectral peaks, compares the target against a compact instrument-profile database, and searches for a playable combination of up to four instruments.

The result is not simply a pitch transcription. SpectraScore asks two connected questions:

1. Which instrument combination best approximates the target's spectral envelope? 2. Which measured spectral components can those instruments play convincingly?

Core idea: orchestration is derived from the acoustical structure of the source. In Spectral voicing, assigned notes come from measured peaks in the actual Sound; they are not restricted to integer harmonics of a single fundamental.

Sound → orchestration

SELECT ONE SOUND ↓ MULTICHANNEL POWER-SPECTRUM ANALYSIS ↓ TARGET DESCRIPTORS centroid / spread / six energy bands spectral root / harmonic confidence ↓ GENERAL MEASURED SPECTRAL PEAKS prominence + adaptive consolidation ↓ INSTRUMENT-COMBINATION SEARCH timbral model + playability + register realism ↓ VOICING ASSIGNMENT unison / octaves / chord / measured peaks ↓ RESULTS Info report MusicXML Strings SpectraScore Results Table optional Picture-window visualization optional .musicxml file

Quick start

Select exactly one Sound object in Praat.
Run SpectraScore.praat.
Leave the default instrument list unchanged to search the full built-in ensemble, or remove abbreviations to restrict the search.
Set Max_combination_size. A value of 4 searches ensembles of one to four instruments in Automatic mode.
Choose Automatic if SpectraScore should decide the ensemble size, or Exact max size if you specifically want that number of instruments.
Choose the instrument spectral profile: pp, mf, or ff.
For acoustically derived pitch assignment, use Spectral (match measured peaks).
Leave Enable_microtones on when you want measured deviations retained in the notation.
Run the analysis. The MusicXML Strings object is selected at the end so it can be passed directly to the AudioTools VST instrument workflow.

Useful starting point: for re-orchestrating a recorded chord, synthesized spectrum, resonant texture, or other complex source, use Spectral voicing with Automatic ensemble size and the mf profile.

Analysis engine

Multichannel spectrum

Every channel is analyzed independently and the channel power spectra are summed. Stereo or multichannel material is therefore not collapsed by waveform summation, avoiding phase cancellation between channels.

Target descriptors

The combined spectrum supplies the descriptors used for orchestration matching:

Descriptor	Role
Spectral centroid	Global brightness / spectral centre of gravity.
Spectral spread	Width of the spectrum around the centroid.
Six-band energy profile	Normalized energy in 50–200, 200–500, 500–1000, 1000–2000, 2000–5000, and 5000–10000 Hz.
Spectral root	A reference pitch found by harmonic summation with direct-frequency support and local refinement.
Root confidence	How clearly the winning spectral root separates from other root candidates.
Odd/even harmonic balance	A weak, confidence-gated harmonic descriptor used only when the Sound is meaningfully explained by the detected root.
HNR / file level	Reported for description and QC; they do not drive the orchestration score.

Root is a reference, not a restriction. In Spectral voicing, the detected root helps describe the Sound and annotate harmonic relationships, but it does not decide which spectral peaks are allowed to become notes.

General measured spectral peaks

Spectral voicing searches for significant local maxima across the measured spectrum rather than only examining H1, H2, H3, and so on. This allows polyphonic and inharmonic components to become orchestration pitches.

The detector combines:

Relative peak floor — rejects components too far below the strongest detected region.
Local prominence — a candidate must rise meaningfully above its local spectral surroundings.
Adaptive consolidation — low-frequency ripples are merged more aggressively than high-frequency peaks.
Valley test — close peaks can remain separate when a real spectral valley supports the interpretation of two distinct components.
Root-anchor rescue — a broad but physically meaningful fundamental can be retained even when it is not a sharply prominent local maximum.

Each selected peak is labelled P1, P2, and so on. If it lies near an integer harmonic of the spectral root, the Info report and Table also show an optional Hn relation.

Example: P1 128.8 Hz [H1] P2 195.8 Hz [non-harmonic] P3 293.4 Hz [non-harmonic] P4 310.9 Hz [non-harmonic] P5 465.8 Hz [non-harmonic]

Why this matters: a recorded chord does not have to behave like one harmonic series. Notes such as D, E-flat or B-flat can be retained as real spectral targets even when they are not integer multiples of the detected root.

Instrument model

The built-in ensemble contains twelve orchestral instruments. The Instruments field accepts the following abbreviations, separated by spaces or commas:

Abbreviation	Instrument	Notation
`BTb`	Bass tuba	Concert pitch
`Bn`	Bassoon	Concert pitch
`Cb`	Contrabass	Written one octave above sounding pitch
`ClBb`	B-flat clarinet	Transposing instrument
`Fl`	Flute	Concert pitch
`Hn`	Horn	Transposing instrument
`Ob`	Oboe	Concert pitch
`Tbn`	Trombone	Concert pitch
`TpC`	Trumpet in C	Concert pitch
`Va`	Viola	Concert pitch
`Vc`	Cello	Concert pitch
`Vn`	Violin	Concert pitch

Two range concepts

Each instrument has an absolute playable range and a narrower preferred / idiomatic range. The absolute range is a hard gate: pitches outside it are rejected. Pitches outside the preferred range remain possible but receive a soft register penalty.

Spectral profile

For pp, mf, and ff, each instrument stores a centroid and spread. SpectraScore converts those values into a six-band Gaussian spectral-envelope model and compares the resulting ensemble profile with the measured target.

Important model limitation: these are compact modeled instrument profiles, not spectra measured from a specific sample library, performer, articulation, register, or room. The result should be read as an orchestration proposal based on spectral tendencies, not as a prediction of the exact spectrum produced by real players.

Ensemble search

Automatic vs exact ensemble size

Mode	Behavior
Automatic	Searches every playable ensemble size from K=1 through Max_combination_size and chooses the lowest final score.
Exact max size	Searches the same sizes for diagnostics, but the selected result is constrained to K=Max_combination_size.

In Spectral voicing, ensemble candidates are evaluated together with their measured-peak assignments. Different instruments preferentially cover different peaks. Doubling a peak is allowed only as a fallback and receives a penalty.

Selection score

Lower scores are better. In Spectral mode the score combines timbral-envelope similarity, global spectral moments, measured-peak coverage, pitch-assignment quality, register realism, duplicate use of peaks, and a small harmonic-balance term.

selection score = 0.32 × band-envelope distance + 0.21 × centroid error + 0.12 × spread error + 0.21 × (1 − measured-peak coverage) + 0.05 × assignment cost + 0.05 × register penalty + 0.03 × duplicate fraction + 0.01 × confidence-gated odd/even error + ensemble-size penalty

The optional ensemble-size penalty is:

size penalty = Additional instrument penalty × (K − 1)

Measured-peak coverage uses square-root-compressed spectral salience rather than raw power. This prevents one dominant fundamental from making a single instrument appear to explain an otherwise rich spectrum.

Voicing strategies

Strategy	Behavior
Unison	Every selected instrument attempts to play the detected spectral root. A root outside an instrument's absolute range makes that assignment unplayable.
Octaves	Places instruments around octave-related versions of the root, with brighter instruments tending upward and darker instruments downward, then folds into playable range.
Chord	Builds a root / third-like ratio / fifth / octave structure according to ensemble size, with octave displacement when needed for range.
Spectral	Assigns instruments directly to significant measured spectral peaks. Peak strength and idiomatic register both influence the assignment.

Spectral is the distinctive mode of the tool: the source spectrum itself becomes the pitch reservoir for the orchestration.

Controls

Main form

Control	Default	Function
Instruments	All 12 instruments	Space- or comma-separated abbreviations. Remove instruments to restrict the search.
Max_combination_size	4	Maximum searched ensemble size; accepted range is 1–4.
Ensemble_size_mode	Automatic	Choose the best K automatically or force K to the maximum size.
Spectral_profile_dynamic	mf	Selects the pp / mf / ff centroid-and-spread profile used for each instrument.
Voicing_strategy	Spectral	Unison, Octaves, Chord, or direct matching to measured spectral peaks.
Enable_microtones	On	Retains quantized cents deviations in MusicXML instead of snapping all notes to equal temperament.
Advanced_settings	Off	Opens the second dialog for search, microtone, and peak-detection controls.
Save_xml_file	Off	Also writes a `.musicxml` file to disk. The in-memory Strings object is created regardless.
Draw_visualization	On	Draws the four-panel SpectraScore report in the Picture window.

Advanced settings

Control	Default	Function
Allow repeated instruments	Off	Allows repeated instrument entries when searching larger combinations.
Additional instrument penalty	0.005	Small complexity cost added for each instrument beyond the first.
Microtone precision cents	12.5	Grid used to quantize measured cents deviations when microtones are enabled.
Peak floor dB	−30 dB	Relative spectral level below which measured peak candidates are ignored.
Min peak separation cents	35 cents	Minimum high-frequency separation; low and mid frequencies use wider adaptive consolidation unless a clear valley supports separate peaks.
Minimum peak prominence dB	3 dB	Minimum local prominence required for a general spectral peak.
Max spectral peaks	24	Maximum number of measured peak candidates retained for Spectral assignment.

Outputs

Output	Purpose
Info report	Target descriptors, measured peaks, best score for each ensemble size, winning instrument combination, and final voicing assignments.
`musicxml_<source>_SpectraScore`	Praat Strings object containing one MusicXML source line per Strings item. It is left selected for direct use with the AudioTools VST instrument host.
`<source>_SpectraScore_Results`	Persistent Table with one row per selected instrument and detailed pitch, score, range, peak, and target/match diagnostics.
`<source>_SpectraScore.musicxml`	Optional disk file when Save_xml_file is enabled.
Picture-window report	Optional visual explanation of the source spectrum, resulting chord, band-energy fit, and ensemble-size decision.

Results Table

The Table is designed for inspection, export, and connection to later AudioTools workflows. Important fields include:

Instrument, Strategy, Voicing
PeakIndex, PeakRel_dB, PeakProminence_dB, RootRelation
SoundingNote, Assignment_Hz, MIDI, Cents, Notated_Hz
WrittenNote and Transpose_semitones
PreferredLow_MIDI, PreferredHigh_MIDI, RegisterPenalty
SelectionScore, AcousticScore, SizePenalty
PeakSalienceCoverage, AssignmentCost, MeanRegisterPenalty, DuplicateFraction
target and matched centroid/spread values plus root frequency and confidence.

MusicXML compatibility: the Strings object follows the AudioTools MusicXML contract — one XML line per String item — so it can be selected directly and rendered through VST_Effect_from_Praat.praat in instrument mode.

Visualization

When Draw_visualization is enabled, the Picture window presents the orchestration decision as a four-part report:

A · Target spectrum

Shows the measured target spectrum on a logarithmic frequency axis, the detected root's harmonic grid, the general measured peak candidates, and the exact frequencies assigned to the selected instruments.

B · Resulting chord

Shows the orchestration on a grand staff in concert pitch. The target root is displayed as a reference, and each instrument note is labelled with its measured frequency, notated frequency, cents value, voicing label, and written pitch where the instrument transposes.

C · Band energy: target vs model

Compares the measured six-band energy profile with the modeled orchestration profile. The target is shown against the contribution of the selected instruments.

D · Ensemble-size decision

Shows the best score found for each searched K. The acoustic component and explicit ensemble-size penalty are visible separately, making it possible to see why one, two, three, or four instruments won.

A summary strip reports the main error terms and reminds the user that the instrument spectra are modeled profiles rather than measured recordings.

Reading the results

Result	How to interpret it
Selection score	Overall search objective. Lower is better. Compare solutions from the same Sound and settings rather than treating it as an absolute perceptual quality scale.
Band-envelope distance	How different the modeled six-band distribution is from the measured target.
Measured-peak coverage	Fraction of compressed spectral salience covered by the assigned distinct peaks. Higher is better.
Assignment cost	Combined cost of choosing strong target peaks while keeping instruments in comfortable registers. Lower is better.
Register penalty	Zero inside the preferred range; rises as an assignment moves toward an instrument's technical extremes.
Duplicate fraction	How much the solution had to reuse the same measured peak for more than one instrument.
Root confidence	Confidence in the spectral reference root. A low value does not invalidate general spectral peaks; it mainly reduces the importance of root-dependent harmonic descriptors.

Do not read “best match” as “the only correct orchestration.” SpectraScore optimizes one explicitly defined acoustic model. The result is best used as a compositional proposal, comparison tool, or starting point for orchestration and audition.

Notes & limitations

Instrument spectra are modeled. The database uses stored centroid/spread values and derived Gaussian band envelopes, not measured samples across articulation, register, player, dynamic, and room.
Equal-energy ensemble model. Instrument profiles are averaged when predicting the orchestration spectrum; individual balance and orchestral mixing are not optimized.
Static analysis. The complete selected Sound is treated as one spectral target. A strongly evolving Sound may be better analyzed in separate snapshots or segments.
Peak assignment is spectral, not perceptual source separation. A measured peak can represent a note, partial, resonance, beating product, or other component. The algorithm does not claim to identify physical sound sources.
Preferred ranges are soft orchestration heuristics. They improve register realism but do not encode every fingering, articulation, dynamic, or performance constraint.
Microtonal notation is quantized to the selected cents grid. The measured peak frequency and the notated frequency are reported separately.
Maximum ensemble size is four. The current search is explicitly designed around compact combinations of up to four parts.
HNR and file level are descriptive. They appear in the report but do not determine the winning orchestration.

Citation

Praat AudioTools: Cohen, S. (2026). Praat AudioTools: An Offline Analysis–Resynthesis Toolkit for Experimental Composition.

Praat AudioTools repository