A primitive parallel-formant speech synthesizer in the browser. Late-70s / early-80s tier (Votrax, SAM).
The name is a portmanteau of Klatt (Dennis Klatt, the formant-synth pioneer) and Klatsch (German for gossip / casual chat).
Commercial support - integration consulting from the author
The klattsch app builds on this engine with a piano-roll editor for singing, pitch editing, multiple voice parts, backing tracks, and WAV and video export. It runs offline on Windows, macOS, Linux, and Android.
klatts.ch · itch.io for desktop · Google Play for Android
You type a phoneme string in ARPABET, with optional directives, and the computer says it.
HH AH L OW hello, default voice
b140 HH AH L OW higher voice
bA3 HH AH L OW higher voice (note name)
AY+15 D IH D "I did" with a rising contour
D IH D DH AE(+40) T "did THAT" with a transient pitch ornament on AE
r200 bC#4 ( HH AH ) ( L OW ) sung syllables, one note per group
See the in-app syntax help panel for the full directive table.
A [voice=N] marker splits the phoneme string into independent voice sections.
Everything before the first marker is voice 0; each marker begins a new section
that compiles from a fresh initial state (running directives never carry across a
marker). The sections are meant to sound together.
bC3 AA r400 AA AA [voice=1] bC4 IY r400 IY IY a two-voice chord
compileString returns a voices array ([{ schedule, totalMs, phrases }])
alongside the backward-compatible top-level schedule/phrases/totalMs, which
mirror voice 0. Render each voice into its own buffer (or its own worklet output)
and mix. With no marker, voices has a single entry and the top-level fields are
unchanged.
npm install klattschThe same package works as a CLI, as an importable engine in Node, and as an embeddable engine + AudioWorklet in the browser. Zero runtime dependencies.
Render a phoneme string straight to a WAV file:
npx klattsch "HH AH L OW" hello.wavimport { compileString, FormantSynth, encodeWav } from 'klattsch';
const sampleRate = 48000;
const { schedule, totalMs } = compileString('HH AH L OW');
const synth = new FormantSynth({ sampleRate, schedule });
const buf = new Float32Array(Math.ceil(totalMs * sampleRate / 1000));
synth.process(buf);
const { bytes } = encodeWav(buf, sampleRate);
// write bytes to a .wav fileimport { compileString } from 'klattsch';
import workletUrl from 'klattsch/formant-worklet.js?url';
const ctx = new AudioContext();
await ctx.audioWorklet.addModule(workletUrl);
const node = new AudioWorkletNode(ctx, 'formant-processor');
node.connect(ctx.destination);
const { schedule } = compileString('HH AH L OW');
node.port.postMessage({ type: 'schedule', schedule });For polyphony, construct the node with one output per voice and address each by a
voice field:
const { voices } = compileString('bC3 AA [voice=1] bC4 IY');
const node = new AudioWorkletNode(ctx, 'formant-processor', {
numberOfOutputs: voices.length,
outputChannelCount: voices.map(() => 1),
});
voices.forEach((v, i) => {
const g = ctx.createGain();
node.connect(g, i, 0);
g.connect(ctx.destination);
node.port.postMessage({ type: 'schedule', schedule: v.schedule, voice: i });
});A reset message with no voice field resets every voice; with one, it resets
just that voice. Offline renders can pass processorOptions: { schedules }.
<script type="module">
import { compileString } from 'https://esm.sh/klattsch';
const ctx = new AudioContext();
await ctx.audioWorklet.addModule('https://esm.sh/klattsch/formant-worklet.js');
const node = new AudioWorkletNode(ctx, 'formant-processor');
node.connect(ctx.destination);
const { schedule } = compileString('HH AH L OW');
node.port.postMessage({ type: 'schedule', schedule });
</script>- Excitation: voiced source is a Rosenberg-style glottal pulse with a tunable open / closed quotient (
g/ "effort") and unvoiced source is xorshift noise. These are crossfaded by each phoneme'svoicingparameter, with optional aspiration noise mixed in. - Filtering: three parallel bandpass biquads for each formant.
- Prosody: the sequencer compiles phoneme strings into a time-stamped schedule of formant targets.
- Voice character: vibrato (depth + rate), aspiration / breathiness, spectral tilt, and glottal effort are all controllable.
The compiler is decoupled from the synthesizer: it produces a schedule of
parameter targets that FormantSynth interprets, but any consumer can interpret
the same schedule differently. Two directives support that:
[engine=NAME]marks a voice section for a named engine. The compiler does not interpret the name; it records it on eachvoices[i].engine(and the top-levelengine) for a consumer to dispatch on.[engine]resets to the default. Compiling with{ engine: 'NAME' }sets the default for every section.- Uppercase bracket directives such as
[OQ=0.65]or[FNZ=450]are recorded as opaque, sticky state that rides into every subsequent schedule target, for an engine that reads extra parameters beyond the built-in set.[OQ]clears the override. Bank fields namedF4/BW4and up are scaled by the runningsscale likeF1-F3.
FormantSynth ignores both, so existing strings and banks are unaffected.
- Klatt, D. H. (1980). Software for a cascade/parallel formant synthesizer.
- Hillenbrand et al. (1995). Acoustic characteristics of American English vowels.
- Rosenberg, A. E. (1971). Effect of glottal pulse shape on the quality of natural vowels.
- Robinson, R. Bristow-Johnson. Audio EQ Cookbook.
- Mokhtari, P. & Tanaka, K. (2000). A Corpus of Japanese Vowel Formant Patterns. Bulletin of the Electrotechnical Laboratory (ETL), Vol. 64, Special Issue, 57-66. (project page, data file) - source of the Japanese vowel formants in the
ja-mokhtari-2000phoneme bank.
- libadlmidi-js - WebAssembly build of libADLMIDI, an OPL3 FM synthesis library with AudioWorklet integration. Where klattsch does parallel-formant vocal-tract synthesis, libadlmidi-js does FM-operator synthesis: the sound of early-80s arcade boards and AdLib cards. Includes oplsfxr, a sfxr-style sound effect generator.
klattsch is built and maintained by Tony Gies. For studios, indie developers, and agencies integrating klattsch into a shipped product, consulting is available through his consultancy, Crash United, LLC.
| Service | Description |
|---|---|
| Game / app integration | Wiring klattsch into your engine (Unity, Godot, web, Electron, Flutter), with dialog-system glue and tooling for non-programmer collaborators (writers, sound designers) |
| Custom character voices | Crafting a recognizable voice signature for a specific character: formant tuning, prosody templates, phoneme calibration, voice tests against scripted dialogue |
| Audio pipeline work | Routing klattsch through your DSP graph: mixing with music, ducking, environmental effects (reverb, distortion, radio filtering), multi-voice ensembles, dynamic vocal sizing |
| Language / phoneme expansion | Non-English phoneme tables, alternate transcription formats, custom symbol sets for stylized worlds (alien races, fantasy languages, in-universe scripts) |
| Performance tuning | Real-time constraints (game audio thread, low-latency targets), WASM/Rust ports, embedded or constrained-runtime targets |
| Custom DSP features | Cascade synthesis, additional formants, LPC pre-filtering, vocoder modes, custom synth extensions beyond the included parallel-resonator engine |
| Priority bug fixes | Reported issues triaged and patched ahead of the public queue, with backports to your pinned version |
| Workshops / talks | Formant synthesis, retro speech tech, or DSP fundamentals for your team |
For pricing, scoping, or anything not listed above, email support@crashunited.com to discuss your project.
To support ongoing development without a formal contract, GitHub Sponsors or Ko-fi are the simplest paths.
MIT © Tony Gies