Write the notes. Shape the breath. Hear the difference between playing a melody and performing a phrase.
Click a note to edit. The curve shows planned pitch, not measured pitch.
One line per note: D4 1 la 0.7 means pitch, beats, syllable, stress. R 0.5 is a rest; _ sustains the preceding syllable on a new pitch. Use one syllable per note; this is not unrestricted text-to-speech.
Explicit phones: [s:t:eh:iy]. Vowels: aa ae ah eh ih iy oh oo uh ax er. Consonants: m n ng l r w y h s z sh zh f v th dh p b t d k g. Best register: roughly C3–C5. Maximum score: 90 seconds.
A periodic source can play notes. A singer-shaped performance also needs phrasing, articulation, connected gestures, and coordinated changes in sound. Cantoria keeps those decisions explicit and inspectable.
Neil Thapen’s open glottal-source implementation supplies the LF waveform parameterization. Cantoria adapts that small component, but uses its own acoustic filter bank, score compiler, rendering pipeline, and performance rules. It does not embed Pink Trombone’s complete vocal-tract solver.
Source and upstream license ↗A working precedent for performative, parametric singing: a source–filter instrument with controllable pitch, breath, tension, and effort. Its official distribution uses CeCILL. Cantoria references the architecture; no Cantor Digitalis code or Max patches are included.
Official project and source links ↗A rules-based articulatory speech system demonstrating the separation of phonetic planning and acoustic production. It is useful prior art for a larger phonetic front end. Its code is not included in this prototype.
GNU’s release announcement ↗| Dimension | A · Straight | B · Performed |
|---|---|---|
| Pitch | Notated targets, no expressive deviation | Connected transitions, onset settling, bounded drift |
| Vibrato | Off | Delayed entry; depth depends on note age |
| Energy | Even level with clean phrase edges | Phrase arc, syllable stress, local swell, release |
| Breath | Same base voice air; no inhalation gesture | Generated inhalations and a falling breath reserve |
| Voice color | Same tract and phoneme targets | Effort-linked source tension and subtle resonance motion |
| Articulation | Same syllables, essential transitions retained | Longer coarticulation and consonant anticipation |
| Fair comparison | Same score length, voice anatomy, sample rate, and dry RMS target. No added backing track. | |
Expression is not just random pitch jitter. One phrase-level control coordinates amplitude, source tension, breath, and resonance. Seeded low-amplitude variation adds texture without replacing that structure. The “breath reserve” is a useful control law—not a solved model of lungs, tissue, and fluid dynamics.
The app computes its own glottal waveforms from equations, moves its resonators using explicit rules, generates its turbulence with a seeded pseudorandom generator, and creates room reflections with feedback delays. There are no voice recordings, pretrained weights, neural networks, speech services, microphone inputs, or external runtime dependencies. The in-memory PCM buffers are newly synthesized outputs, not sample-library inputs.
Audio is rendered in a Web Worker, then played through Web Audio. Editing a synthesis parameter rerenders the score; this is a score-to-performance instrument, not a low-latency live vocal controller. A/B and room levels switch during playback. A same-thread fallback is available when a browser blocks blob workers.
Cantoria is © 2026 Collin Christensen, all rights reserved; see the LICENSE of the site that hosts this page. The adapted Pink Trombone component retains its upstream notice below. # Cantoria — third-party notices The LF glottal-waveform parameterization in src/engine.js (lfParameters and lfValue) is adapted from Pink Trombone, originally by Neil Thapen, as exposed in the TypeScript modularization by chdh: https://github.com/chdh/pink-trombone-mod/blob/master/src/Glottis.ts https://github.com/chdh/pink-trombone-mod/blob/master/LICENSE.md Only that source component is adapted. The complete Pink Trombone waveguide, UI, noise generator, and controller are not bundled. Cantoria uses its own source/filter renderer and performance planner. The attribution does not imply endorsement by the upstream authors. Upstream license (retained in full): MIT License Copyright 2017 Neil Thapen Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. Research references, not included dependencies: Cantor Digitalis: https://chorus-digitalis.lam.jussieu.fr/cantor-digitalis.html The official source distribution uses CeCILL. No code or patches copied here. Gnuspeech: https://lists.gnu.org/archive/html/info-gnu/2015-10/msg00007.html No Gnuspeech code or language resources are included. No external audio, images, typefaces, JavaScript packages, model weights, phonetic corpora, or prerecorded impulse responses are embedded.
The buttons below run fresh engineering checks in your browser. They establish that the synthesis works and that the performance layer changes it. They do not establish that the result is perceptually indistinguishable from a human singer.
Start with Vowel aria, Room at Dry, and switch A/B on a held note. Then use Afterglow to judge whether you can understand the words without looking at the score. Finally, set Expression to zero: the two modes should converge to identical output. That last comparison checks that the performance layer really is removable.
For a stronger study, randomize and blind A/B presentation, retain RMS matching, use multiple listeners and previously unheard melodies, and rate naturalness, expressive intention, and lyric intelligibility separately. Add genuine human recordings only as evaluation references, not as synthesis inputs. Do not equate a waveform change or a low pitch error with a human-quality singing voice.
The next useful work is not adding more random wobble. It is fitting richer rule-based coarticulation, adding nasal zeros and side branches, improving glottal register transitions, expanding stress and syllable planning, and tuning against blinded listening results. The renderer’s explicit control traces make each intervention independently testable. A more complete articulatory backend can replace the current tract filters without discarding the performance planner.