CHANTER

a python speech synthesizer inspired by IRCAM's CHANT system—FOF (fonction d'onde formantique) grains shaped by coarticulation and phrase dynamics.

what it does

transforms text into wavering, vowel-rich audio via:

• FOF grain synthesis — each glottal pulse spawns decaying sinusoids at formant frequencies
• coarticulation engine — phonemes influence neighbors, smoothing transitions like actual vocal tracts
• FORMES-style scheduling — phrase contours, stress patterns, final lengthening emerge from rule systems
• register modeling — formant tuning shifts with pitch, simulating soprano/alto/tenor/bass resonance
• jitter & shimmer — multi-rate random wandering keeps it organic

Architecture

Text
↓
NLPParser
↓
FORMES (phrase + phoneme rules)
↓
Sparsegram (time-indexed control frames)
↓
CHANTEngine
↓
.wav

quick start

python CHANTER.py "hello world" -v tenor -o hello.wav

python CHANTER.py "what is the sound of one hand clapping?" -v alto --speed 0.5

arguments

flag	default	meaning
-o/--out	auto	output wav path
-f/--pitch	voice-dependent	base f0 in hz
-s/--speed	0.38	phoneme duration scale
-v/--voice	tenor	soprano/alto/tenor/bass
-t/--tempo	1.0	global tempo multiplier
-d/--dynamics	0.7	effort/loudness 0-1
--lookahead	8	coarticulation window
--phonemes	built-in	custom yaml/json phoneme table
--text-file	—	read input from file

dependencies

numpy scipy pyyaml(optional)

phoneme customization

drop a phonemes.yaml alongside the script or pass --phonemes path/to/custom.yaml:

AA:
  phoneme_type: vowel
  formants: [[700, 130, 0], [1220, 70, -6], [2600, 160, -24]]
  voicing: 1.0
  glottal_tension: 0.5

each formant tuple: [center_freq_hz, bandwidth_hz, amplitude_db]

how it works

1. NLPParser splits text on punctuation into breath groups, maps letters→phonemes via digraph+simple lookup
2. FORMES scheduler applies phrase dynamics, stress accents, legato rules
3. CoarticulationEngine blends formant targets across phoneme boundaries
4. RegisterModel shifts formants based on f0 and voice type
5. CHANTEngine renders: for each glottal period, spawns FOF grains per formant, applies antiformants, mixes noise for fricatives/plosives

lineage

inspired by xavier rodet's CHANT (1984) and the FORMES composition environment—where synthesis parameters became musical gestures and vocal models sang ircam's electroacoustic dreams.

What is Chant?

The CHANT Project: From the Synthesis of the Singing Voice to Synthesis in General

to do

1. add streaming
2. reduce compute costs
3. some of the frequency sweeps could be manually tightened up
4. midi?

license

do what thou wilt