8-Channel Speech-Driven Spatialization — User Guide

Maps detected pitch to azimuth around an octophonic speaker ring and maps intensity to a proximity-gain control. The active direction is rendered with adjacent-speaker constant-power panning, supported by a low coherent all-speaker bed.

Author: Shai Cohen Affiliation: Department of Music, Bar-Ilan University, Israel Version: 0.6.1 (2026) License: MIT License Category: Spatial Audio / Analysis-Driven Processing Repo: Praat AudioTools

Contents:

What this does Quick start Speaker geometry Analysis pipeline Pitch → azimuth Intensity → proximity gain Panning & ambient bed Presets Parameters Output formats Visualization Implementation notes

What this does

The script analyzes one source Sound frame by frame. Pitch controls angular position around a fixed eight-speaker ring, while intensity controls an amplitude multiplier called proximity gain. At every control frame the pitch-derived angle is panned between the two adjacent loudspeakers. A user-controlled coherent bed is added beneath the directional pan, all eight gains are renormalized together, and the proximity gain scales the resulting field.

What is “speech-driven” here?

The control signals are Praat pitch and intensity analyses, which make the tool particularly natural for speech and voice. The script does not classify phonemes, words, prosody categories, or linguistic content. Any source can be processed; if no voiced pitch is detected, the azimuth is held at the centre of the selected pitch arc while intensity continues to control gain.

Pitch → azimuth: fixed floor/ceiling mapping by default, with an optional adaptive per-file mapping.
Intensity → proximity gain: normalized between the file's 5th and 95th intensity percentiles.
Adjacent-speaker pan: cosine/sine constant-power law between neighboring speakers.
Coherent bed: the same signal is present at low level in all eight speakers before renormalization.
Five output layouts: octophonic, opposing stereo stems, two quad groups, four-channel fold-down, or stereo fold-down.
Suite-style visualization: octagon trajectory, pitch contour, proximity-gain contour, opposing output examples, and summary.

Quick start

Select exactly one Sound in Praat.
Run 8-Channel_Speech-Driven_Spatialization.praat.
Choose a Preset. The default is Custom.
For consistent pitch-to-position behavior across recordings, leave Pitch_mapping = Fixed.
Adjust Min_proximity_gain, Max_proximity_gain, and Ambient_level if needed.
Choose the required Output_format.
Leave Draw_visualization on to inspect the mapping and resulting motion.
Click OK.

Default mapping: 75 Hz maps to 225° and 600 Hz maps clockwise to 45°. Intensity maps from 0.2 to 1.0 proximity gain, using the file's 5th–95th percentile intensity range. The default coherent bed is 0.01.

Speaker geometry

The listener is at the origin. The speaker ring uses y upward/front, x rightward, angle 0° at Front, and increasing angle clockwise.

Channel	Label	Angle	Cartesian role
1	FL	315°	Front-left diagonal
2	F	0°	Front
3	FR	45°	Front-right diagonal
4	R	90°	Right
5	BR	135°	Back-right diagonal
6	B	180°	Back
7	BL	225°	Back-left diagonal
8	L	270°	Left

For speaker angle a: x = sin(a) y = cos(a)

The same speaker coordinates are used by the octophonic renderer, the geometric stereo fold-down, the four-channel fold-down, and the visualization.

Analysis pipeline

Working source

Exactly one Sound is required. Multichannel input is converted to mono before analysis and rendering. If the selected Sound has a non-zero start time, the working copy is re-extracted so that its analysis and AmplitudeTier domains begin at time 0.

Multichannel note: the mono conversion is Praat's ordinary channel averaging. There is no cancellation-safe strongest-channel fallback, so strongly anti-correlated channels can reduce the analysis signal.

Pitch

To Pitch: time step = Time_step floor = Pitch_floor ceiling = Pitch_ceiling

The script also measures the file's observed pitch minimum, maximum, and mean. If no voiced frames exist, those statistics are undefined; the script then substitutes the centre of the requested pitch range and uses a fixed centre-of-arc azimuth.

Intensity

To Intensity: minimum periodicity = Intensity_floor time step = Time_step If Intensity_floor ≤ 0: minimum periodicity = Pitch_floor effective analysis window ≈ 3.2 / minimum_periodicity

Intensity is normalized using the 5th and 95th percentiles rather than the absolute minimum and maximum. If the percentile range is unavailable or degenerate, the script falls back to the raw extrema.

Control frames

frameShift = Time_step numberOfFrames = floor(duration / frameShift)

The source must be at least one time step long. Pitch and intensity are sampled at each control frame. The resulting channel gains are written to eight AmplitudeTiers, which interpolate linearly between those points.

Pitch → azimuth

Fixed mapping — default

Fixed mapping normalizes pitch against the explicit Pitch_floor and Pitch_ceiling. Values outside the range are clipped. This means the same pitch is mapped to the same azimuth in different files.

pitchNorm = (pitch - Pitch_floor) / (Pitch_ceiling - Pitch_floor) pitchNorm is clipped to 0 ... 1 angleRange = clockwise arc from Low_pitch_angle to High_pitch_angle targetAngle = Low_pitch_angle + pitchNorm × angleRange

Adaptive mapping

Adaptive mode uses the observed minimum and maximum pitch of the current file. Whatever pitch range that recording contains is stretched across the complete selected angular arc. It therefore emphasizes within-file contour rather than preserving a stable pitch-to-angle calibration across sources.

pitchNorm = (pitch - observed_minimum) / (observed_maximum - observed_minimum)

Unvoiced frames

After a valid pitch has appeared, unvoiced frames retain the last valid pitch. Before the first voiced frame, there is no previous value to hold, so the file's mean detected pitch is used. If the entire file has no voiced frame, the position is held at the midpoint of the angular arc.

Angular direction: mapping always follows the clockwise arc from the low-pitch angle to the high-pitch angle. With the default 225° → 45° mapping, the 180° arc passes through 270°, 315°, 0°, then 45°.

Intensity → proximity gain

Intensity is converted to a normalized value using the 5th–95th percentile range, then interpolated between the minimum and maximum proximity-gain settings.

intensityNorm = (intensity - P05) / (P95 - P05) intensityNorm is clipped to 0 ... 1 proximityGain = Min_proximity_gain + intensityNorm × (Max_proximity_gain - Min_proximity_gain)

What does “proximity gain” mean?

It is an amplitude cue only. The script contains no metres, inverse-distance law, propagation delay, air attenuation, early/late reflection model, or direct-to-reverberant ratio. The visualization draws louder states nearer the listener as a visual metaphor for this gain, not as a physical source-distance calculation.

The source already contains its original dynamics. Multiplying it by a gain derived from the same intensity therefore tends to expand dynamic contrast: louder moments receive a larger multiplier and quieter moments receive a smaller multiplier. Narrowing the proximity-gain range reduces this effect.

Equal values: if Max_proximity_gain is entered less than or equal to Min_proximity_gain, the script internally sets it to Min + 0.001. This makes the gain variation extremely small, but not mathematically zero.

Panning & coherent ambient bed

Adjacent-speaker selection

For each target azimuth, the nearest loudspeaker is found by circular angular distance. The sign of the wrapped angular offset then selects the neighboring speaker on the correct side, including the 0°/360° boundary.

Constant-power pair

p = angular distance from nearest speaker / angular distance between the two adjacent speakers θ = p × π/2 main gain = cos(θ) adjacent gain = sin(θ) main² + adjacent² = 1

The sine/cosine law reaches the endpoints smoothly. This replaces a square-root law whose slope became singular at exact speaker positions.

Ambient bed

Ambient_level is used as a base value on all eight channels. The two active pan gains are added on top of that base. The complete eight-channel vector is then normalized together before proximity gain is applied.

for every channel i: raw[i] = Ambient_level raw[main] += cos(θ) raw[adjacent] += sin(θ) norm = sqrt(Σ raw[i]²) gain[i] = raw[i] / norm × proximityGain

At each control frame this guarantees:

Σ gain[i]² = proximityGain²

The bed is coherent, not diffuse. It consists of synchronous copies of the same mono source. There is no decorrelation, delay, reverb, random phase, or diffuse-field simulation.

Between control frames

The constant-power identity is exact at the calculated control frames. Praat then linearly interpolates each channel's AmplitudeTier independently, so the summed power can dip slightly between frames. Faster movement therefore benefits from a smaller Time_step. The script report gives reference examples of approximately -0.03 dB at 4.5° per frame and -0.67 dB at 22.5° per frame.

Presets

Presets replace only the pitch floor, pitch ceiling, and low/high angular endpoints shown below. Other controls—including Fixed/Adaptive mapping, proximity gain, ambient level, time step, and output format—remain as set in the form.

Preset	Pitch range	Low pitch angle	High pitch angle	Effect
Custom	Form values	Form value	Form value	No pitch fields are overridden.
Full Range	75–600 Hz	225°	45°	Broad default speech/voice range.
Voice Range	100–300 Hz	225°	45°	Narrower pitch calibration.
Narrow Range	150–250 Hz	225°	45°	Expands a narrow band across the selected arc in Fixed mode.
Extended Range	50–800 Hz	225°	45°	Wider detector and mapping range.
Inverted	75–600 Hz	45°	225°	Reverses the default pitch-to-space direction.

Parameters

Parameter	Default	Actual role
Preset	Custom	Selects one of the pitch-range/angle mappings above.
Pitch_floor	75 Hz	Pitch detector floor and Fixed-mode mapping minimum.
Pitch_ceiling	600 Hz	Pitch detector ceiling and Fixed-mode mapping maximum; must exceed Pitch_floor.
Pitch_mapping	Fixed	Fixed floor/ceiling calibration or Adaptive observed-range mapping.
Low_pitch_angle	225°	Azimuth assigned to normalized pitch 0.
High_pitch_angle	45°	Azimuth assigned to normalized pitch 1, following the clockwise arc.
Min_proximity_gain	0.2	Gain at/below the 5th-percentile intensity; negative values are clamped to 0.
Max_proximity_gain	1.0	Gain at/above the 95th-percentile intensity. If ≤ minimum, it becomes minimum + 0.001.
Ambient_level	0.01	Coherent base added to all eight channels before renormalization; clamped to 0–1.
Time_step	0.01 s	Pitch/intensity sampling interval and spatial-control interval.
Intensity_floor	0	Minimum periodicity for Intensity analysis. 0 or less follows Pitch_floor.
Output_format	8-channel octophonic	Selects the returned routing layout.
Scale_peak	0.95	Target of the shared eight-channel gain stage; invalid values reset to 0.95.
Draw_visualization	on	Draws the v0.6.1 analysis/process figure.
Play_result	on	Plays the output, or a temporary geometric stereo monitor for multi-object stem formats.

Output formats

Shared gain before routing

After the eight AmplitudeTier-processed channels are created, the largest absolute peak across all eight is measured. Unless all channels are silent, one common gain is applied to every channel:

peakAll = max(peak Ch1 ... peak Ch8) sharedGain = Scale_peak / peakAll all Ch1 ... Ch8 *= sharedGain

This preserves inter-channel level relationships. The factor may attenuate or boost the complete bank. If the bank is effectively silent, the normalization stage is skipped.

Format	Returned objects	Routing	Second normalization
8 channels — octophonic	1 × 8-channel	out1–out8 = Ch1–Ch8	No
4 opposing stereo pairs	4 × stereo	Ch1\|Ch5, Ch2\|Ch6, Ch3\|Ch7, Ch4\|Ch8	No
2 quadraphonic groups	2 × 4-channel	Diagonal = Ch1,3,5,7; Cardinal = Ch2,4,6,8	No
4-channel fold-down	1 × 4-channel	FL, FR, BR, BL with cardinals split equally to neighboring corners	Peak-scaled to Scale_peak
Stereo fold-down	1 × stereo	Every speaker is equal-power panned from its x coordinate	Peak-scaled to Scale_peak

Four-channel fold-down

FL = Ch1 + 0.7071 × (Ch2 + Ch8) FR = Ch3 + 0.7071 × (Ch2 + Ch4) BR = Ch5 + 0.7071 × (Ch4 + Ch6) BL = Ch7 + 0.7071 × (Ch6 + Ch8)

Stereo fold-down

For each speaker, its x coordinate determines a constant-power stereo position:

p = speaker_x, clipped to -1 ... +1 a = (p + 1) × π/4 L = cos(a) R = sin(a)

This preserves the left/right geometry of the octagon but necessarily discards front/back distinction.

Playback of stem formats

When the selected format returns four stereo pairs or two quad groups, the script creates a temporary stereo preview using the geometric stereo fold-down above, peak-scales it, plays it, and removes it. The preview is not one of the returned output objects.

Visualization

The current figure uses an 8 × 8 canvas and reads the same pitch, intensity/proximity, and working output channels used by the DSP.

TITLE • source, preset • pitch-analysis range • Fixed / Adaptive mapping • pitch-angle endpoints • output format PANEL A — Speaker map & path • octophonic speaker ring and listener • pitch controls trajectory angle • displayed radius follows proximity gain: high gain is drawn nearer the listener low gain is drawn farther away • this radius is a visual amplitude cue, not physical distance • start point shown separately PANEL B — Pitch contour • time-domain pitch trace • Fixed mode uses Pitch_floor / Pitch_ceiling as plot range • Adaptive mode uses the file's observed range • dotted lines show the range used by the mapping PANEL C — Proximity gain • gain trace between Min_proximity_gain and Max_proximity_gain • dotted reference lines at the configured limits • explicitly labelled as amplitude, not distance PANEL D — Opposing output examples • blue = Ch2 Front • orange = Ch6 Back • one shared amplitude scale • drawn from the working spatial channels before output-format cleanup PANEL E — Summary • source, preset, duration • frame count and Time_step • intensity analysis floor • pitch and intensity ranges • coherent-bed setting • output format, object count, channel count and routing

Trajectory radius: the figure uses a normalized visual radius from approximately 0.88 at minimum proximity gain to 0.38 at maximum gain. That radius has no role in the audio panning calculation.

Implementation notes & limitations

Version label: the script's top header still says 0.6, while its v0.6.1 changelog, Info report, and visualization fixes identify the current implementation as v0.6.1. This guide follows the latter.
One source only: the script checks the selection before displaying its form.
Source duration: sources shorter than one Time_step are rejected.
Mono-derived field: original multichannel spatial information is not preserved.
No voiced pitch: the angular position remains at the arc centre; intensity can still modulate proximity gain.
Percentiles are robust, not immune to silence: they reduce sensitivity to isolated extremes, but enough silent material can still enter the lower percentile.
Intensity is reused as gain: proximity gain deliberately exaggerates source dynamics because it multiplies audio whose amplitude already reflects those dynamics.
Bed semantics: the current DSP places the coherent base on all eight channels, including the two active channels, then adds the directional pan and renormalizes. An older Info sentence still calls it a bed on the six non-active channels; this guide documents the actual v0.6.1 code.
Constant power: exact at analysis/control frames; slightly approximate between frames because the eight AmplitudeTiers interpolate independently.
No acoustic distance simulation: no propagation delay, inverse-distance attenuation, HRTF, room model, or reverberation is calculated.