Frozen Canon

Turning a single melodic phrase into an accumulating harmonic field: the original melody continues while its notes remain as sustained, independently routed voices.

Author: Shai Cohen Affiliation: Department of Music, Bar-Ilan University, Israel Library: Praat AudioTools Category: CDP Version: 0.6 (2026) Script license: MIT CDP engines: LGPL-2.1-or-later

The musical idea: melody becomes harmony

A melody normally moves from one note to the next: each new pitch replaces the previous one. Frozen Canon changes this relationship. It selects stable moments from a recorded melodic phrase, extends each moment using the Composers Desktop Project's Freeze process, and lets the frozen tones accumulate as the original phrase progresses.

The source melody remains audible in the result. Each selected pitch also becomes a sustained layer, entering at the time of its original event and continuing into a shared, slowly decaying tail. The result is a kind of self-derived harmonic resonance: a sequence of pitches becomes a simultaneous texture without imposing a new chord progression.

The source is both the melody heard in the foreground and the material from which its own harmonic afterimage is constructed.

The frozen notes are not transposed. Their pitch identities come from the original phrase; the CDP Freeze engine iterates short time-domain excerpts rather than constructing a static spectral frame. Depending on the recording, individual voices may retain breath, vibrato, and other local fluctuations.

How the instrument works

The entire workflow is launched from one Praat script. Praat performs event detection, timing, routing, mixing, decay, and the optional Picture visualization. The bundled CDP engine is invoked for each individual frozen voice. No Python is required.

Event selection

Find notes, not random slices

Automatic analysis looks for pitch changes and voiced restarts, then checks short-term pitch stability. A selected note's freeze excerpt is taken slightly after the event onset, rather than treating a noisy attack as the sustained pitch. When there are more candidates than the chosen voice limit, an ordered subset is selected across the phrase.

Manual alternative

PointProcess as a score

Optionally select a Sound and a PointProcess together. PointProcess times provide onset candidates in the Sound's original time domain. The script still needs a reliable local pitch estimate; a mark alone does not guarantee that a frozen voice can be generated.

Freeze is not spectral locking. CDP Freeze fluidly iterates a selected waveform excerpt. The engine's pitch-shift and amplitude-cut settings are held at zero; each voice has a fixed, reproducible random seed and a short entrance fade.

From melody to harmonic field

The original is heard during the opening phrase; the sustained notes are heard alongside it. As each voice enters, more of the source's melodic history becomes present at once. At the end of the phrase the source stops, while the combined frozen texture continues to decay.

Illustrative accumulation timeline (not a measured render) 0 s510152025 Channel 1Channel 2Channel 3Channel 4 Original phrase + accumulating voicesShared frozen-cloud decay (1.5 × phrase)
Original present in each channelIndependent frozen voiceShared tail
The diagram uses a hypothetical 10-second phrase and four selected notes. In the actual result, event times and the number of channels come from analysis or supplied PointProcess marks.

One decay, not a sequence of individual envelopes

All frozen voices remain at their rendered levels until the source phrase ends. Their summed cloud then follows the same exponential gain curve to approximately -60 dB, with a very short final fade to silence. The source phrase is not subjected to the cloud decay, and its amplitude envelope is not reapplied to the frozen notes.

Source duration = T
Cloud decay duration = T × Cloud_decay_factor
Total duration = T × (1 + Cloud_decay_factor)
For the default factor 1.5: 10 s source + 15 s decay = 25 s result.

Multichannel and stereo outputs

The Output_format setting changes the arrangement of already computed frozen voices, without altering the source-based pitch material.

Default · N channels

Multichannel accumulation

Each selected frozen voice occupies one dedicated channel. The original phrase is present in every channel, in the same position on the timeline and with the same original pitch. The number of output channels equals the number of voices actually selected, with a maximum of 16. Thus, five detected and selected voices generate a five-channel Praat Sound.

Alternative · 2 channels

Stereo harmonic field

The same frozen voices are panned across stereo from left to right according to their entry order using equal-power panning. The mono working copy of the original phrase is shared centrally across the stereo output. The stereo result always has two channels, regardless of the number of frozen voices.

Multichannel listening: The multichannel Sound is a discrete set of stems, not an automatic surround-speaker layout. Playback and monitoring depend on the audio interface and host configuration. For an ordinary two-channel listening test, choose Stereo.

If the selected source has two or more channels, the script first makes a mono working copy for pitch detection, freeze processing, and the common original layer. The input Sound object remains unchanged, but its original multichannel spatial image is not preserved in the rendered canon.

Getting started

Installation: Save Frozen_Canon.praat as an independent script in plugin_AudioTools/CDP/. It uses internal/cdp_engine.praat, CDP_Processes.tsv, CDP_Parameters.tsv, and the platform's bundled CDP executables. The supplied Windows engine is supported through the CDP integration; other systems require a compatible bundled engine. No standalone CDP installation, SoundLoom, or Python is required when the bundled engine for the platform is available.
  1. Select one Sound in Praat. A clear monophonic melody with distinct sustained notes is the most straightforward starting point. The script accepts material of at least 0.35 seconds.
  2. Optional: also select one PointProcess. Use it to specify note-onset candidates instead of automatic pitch-jump detection.
  3. Run Frozen_Canon.praat. Start with Custom, the baseline configuration, or Harmonic Accumulation, which uses the same principal settings.
  4. Choose Output format. Select Multichannel for one frozen note per channel, or Stereo for an immediately monitorable left-to-right field.
  5. Render and listen. The result appears as a new Praat Sound; the selected original remains intact. With Draw_visualization on, the Picture window also presents the actual event map, routing, decay, and channel waveforms.
  6. Compare presets using the same source. Use wider event spacing for a sparse harmonic trail or a longer cloud tail for extended resonance.
File placement matters: the script resolves the CDP engine relative to its own CDP/ directory. Do not move it to another category without updating the include and registry lookup paths.

Compositional presets

Presets change the number of selected voices, the common decay length, and the balance between the source and the frozen cloud. They do not introduce transposition or change the core one-voice-per-channel routing rule.

PresetMaximum voicesDecay factorFreeze randomnessOriginal levelFrozen level
Custom51.50.080.850.65
Transparent Harmonic Trail40.80.020.900.40
Harmonic Accumulation51.50.080.850.65
Dense Harmonic Cloud81.50.060.800.65
Long Resonance62.50.030.800.60
Sparse Canon31.50.030.850.70

Transparent Harmonic Trail

Four possible voices, a shorter tail, and a quieter frozen layer let the source melody remain more prominent.

Harmonic Accumulation

The baseline relationship: five potential voices with a cloud tail lasting one and a half times the phrase.

Dense Harmonic Cloud

Up to eight voices; the script also lowers the minimum event spacing to 0.15 s and the pitch-jump threshold to 1.2 semitones.

Long Resonance

A tail lasting 2.5 times the source duration. For a 10-second phrase, total output duration becomes 35 seconds.

Sparse Canon

Up to three voices; automatic detection uses 0.40 s minimum spacing and a 2.5-semitone jump threshold.

Custom

Exposes the baseline parameters for direct experimentation. The preset uses the form's current values rather than a separate parameter file.

A preset sets a maximum number of voices, not a guaranteed number. Fewer notes may be rendered when the source offers fewer usable events. The descriptive effects above are compositional intentions, not guaranteed outcomes for all recordings.

Controls and output options

ControlFunction and constraints
PresetCustom plus five predefined compositional settings.
Maximum_voicesUpper bound on frozen voices; at most 16. Output channel count in Multichannel mode equals the actual number selected.
Cloud_decay_factorRatio of the post-phrase cloud decay to source duration. Default 1.5; maximum 4. Output duration is capped at 180 s.
Pitch_floor_Hz, Pitch_ceiling_HzRange for pitch-based event analysis; defaults 140 and 1500 Hz. The ceiling must exceed the floor and stay below 45% of the sampling rate.
Min_event_gap_sMinimum spacing between automatically selected event candidates; default 0.22 s. Dense and Sparse presets override it.
Pitch_jump_semitonesThreshold for detecting melodic changes; default 1.5 semitones. Dense and Sparse presets override it.
Freeze_window_msRequested note-excerpt size, default 160 ms; valid range 70-600 ms. Excerpts are constrained by nearby event and source boundaries.
Freeze_delay_msNominal CDP iteration delay, default 45 ms; the script limits it relative to the selected freeze region.
Freeze_randomnessCDP iteration delay randomness, 0-1; default 0.08. Pitch shift and CDP amplitude cut remain zero.
Voice_fade_in_msShort entrance fade for each frozen voice, default 45 ms.
Original_level, Frozen_levelRelative source and cloud gains, defaults 0.85 and 0.65. Actual frozen voice gain also includes local RMS balancing and a fixed headroom factor.
Output_formatMultichannel (default) or Stereo; formats are mutually exclusive per run.
Draw_visualizationShows the compositional picture without altering audio; enabled by default.
Play_resultPlays the rendered Sound after processing; enabled by default. For discrete multichannel material, playback compatibility depends on the system.

Result objects

Output formatPraat Sound suffixChannels
Multichannel_FrozenCanon_CDP_MCNumber of selected frozen voices, 1-16
Stereo_FrozenCanon_CDP_Stereo2

The renderer performs a downward-only final peak adjustment if the result exceeds 0.95; it does not boost quiet material to a target peak.

Reading the compositional visualization

The Draw_visualization control produces an optional Praat Picture, drawn after the audio has been rendered. It reports the selected events and the actual output rather than modifying audio.

Panel A

Phrase and events

The source waveform, rendered onset positions, approximate fundamental frequencies, and highlighted source excerpts used for freezing. Other detected candidates may appear as muted markers.

Panel B

Accumulation timeline

The common original-phrase interval and the entry and duration of each frozen voice. Multichannel labels describe dedicated output channels; stereo labels indicate each voice's pan position.

Panel C

Cloud decay

The decay envelope used by the actual frozen layer, starting when the source ends and approaching -60 dB at the output boundary.

Panel D

Rendered output channels

Waveforms measured from the new multichannel or stereo Sound, with a common amplitude reference where compared.

The summary strip reports the source and output durations, detected voices, routing, decay factor, and output level. For many channels, the Picture may abbreviate lane labeling to remain legible; the Sound retains its full channel count.

Technical and listening notes

Source-dependent harmony

The system does not quantize pitches to a scale or choose harmonic intervals. The harmonic field follows the original note collection, including its intonation and performance irregularities.

Pitch tracking limits

Rapid ornamentation, noisy attacks, weak fundamentals, and multiple simultaneous pitches may confuse automatic note detection. A manually marked PointProcess can make onset selection more deliberate, while pitch still has to be trackable.

Time-domain persistence

CDP Freeze creates fluid repeated material; sustained layers may contain audible internal cycles, beating, or periodic motion. The script does not claim perfectly stationary spectral synthesis.

Room-like, not a modeled room

Frozen Canon can sound like a reverberant afterimage of the melody. It is not convolution reverb, an impulse-response model, or conventional feedback delay. Its resonance emerges from sustained notes drawn from the source itself.

Scope of verification: This page documents the supplied v0.6 script's controls and logic. Sound quality and the number of usable frozen voices depend on the input material and on the bundled CDP engine available on the running platform.