Sound Modeling framework using the Web Audio API
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

Steller is a small framework for building composeable sound models in Javascript using the Web Audio API. It features a GraphNode abstraction that encapsulates dynamic signal flow graphs, Param objects to add parameters of various kinds to objects and a Scheduler for sequencing audio events just in time. These classes are all available under the namespace org.anclab.steller.

Examples -

  1. A "live coding" style playground for trying out Steller.
  2. Tala Player is a free player for the various Carnatic tala patterns. The visualization and the sound are both expressed using a single steller composition.
  3. See Gurur Brahma for a fairly sophisticated example of using Steller to present a simple chant melody with precisely synchronized animations.


Using Bower

You can install the latest requirejs version of steller as a Bower component. This is the preferred method now due to its ease.

bower install steller

Using RequireJS

If you already use RequireJS, the AMD release of steller is likely the easiest starting point for you.

<script src="//"></script>
<script src=""></script>
    require(["steller"], function (steller) {
        // Make some noise.
        var sh = new steller.Scheduler(new steller.AudioContext);
        var ch = sh.models.chime().connect();[72,76,79].map(function (p) { return ch.note(p, 0.5, 0.5); })));

You can host your own copy of the steller_amd.min.js file too.



Encapsulates a sub-graph between a set of input nodes and a set of output nodes. The two primary methods it adds are -

  • connect(targetGraphNode, outPinIx, inPinIx)
  • disconnect(targetGraphNode, outPinIx)

Usage: var obj = steller.GraphNode(obj, [in1, in2, ...], [out1, out2, ...]) where obj is an object to turn into a GraphNode, inX are input nodes and outY are output nodes (either AudioNode objects or GraphNodes).


Param objects encapsulate dynamic parameters to sound models and provide some basic glue functionality such as watching for changes and binding to HTML UI elements.

Usage: steller.Param(spec) creates a Param object.

  • obj.param = Param({min: 0, max: 100, value: 50})
  • obj.param = Param({min: 0, max: 100, audioParam: anAudioParam})
  • obj.param = Param({options: ["one", "two", "three"], value: "one"})
  • (value, param) {...});
  • obj.param.unwatch(aPreviouslyInstalledCallback)
  • obj.param.changed() forces observer notification.
  • obj.param.bind(elem, optional<Scheduler>)
  • obj.param.unbind(elem)
  • Param.names(obj) gets names of all exposed params in obj


obj.griffinStrength = Param({min: 0, max: 100, value: 20});
obj.griffinStrength.value = 40;
console.log(obj.griffinStrength.valueOf()); // Prints 40.
console.log(obj.griffinStrength.value);     // Prints 40.
obj.griffinStrength.bind('#strength'); // Binds to a slider described using a document.querySelector string.


A sound model is just an alias for a graph node. Ok not quite. It is a grpah node, but with the connect and disconnect calls observable.

This is available as steller.SoundModel.

Usage: var sm = steller.SoundModel(obj, [inputs..], [outputs..]);

  • sm.on('connect', function () { console.log('connected'); })
  • sm.on('disconnect', function () { console.log('disconnected'); })


A just-in-time scheduler for continuous temporal behaviour. The Web Audio API features sample accurate scheduling of triggered sounds, oscillators and parameter curves. For interactive music applications, these need to be specified just in time so that the system feels responsive to user input, while not sacrificing timing precision. The Scheduler lets you focus on specifying the relationships between the parts of your composition (which can include sound as well as visuals) while taking care of the performance of the specification.

Scheduler threads a clock object through sequential events part of a single "track". Multiple tracks can be spawned or forked in parallel and each can have its own clock object. The clock object tracks both absolute time and a rate integrated pseudo time that can be used for things like tempo changes. The complexity of the scheduler is proportional to the number of simultaneously running tracks.

Usage: var sh = new steller.Scheduler(audioContext); Use the various methods of sh to create a specification or "model" of the desired performance, and then to play it.

Here is a bare bones demo of using the scheduler. To try out the sample code below, open that demo page. Once the page loads, open the developer console. You can now copy-paste the code into the console to run it.

You use the methods of the scheduler object to make specifications or "models" and call to play a built specification.

  • sh.audioContext gives the audio context whose time base is used for scheduling.
  • sh.cancel() cancels all scheduled processes.
  • causes the given model to be played now.
  • sh.stop is a model that will cause the track in which it occurs to terminate.
  • sh.cont is a model that is a no-op when encountered in a track.
  • (clock) {...}) will cause the given callback to be called at the time the model is performed. The resultant action itself has zero duration. Here is a "ping" model -
var ping = function (freq) {
    var decayTime = 0.5 * 440.0 / freq;
    return (clock) {
        var osc = sh.audioContext.createOscillator();
        osc.frequency.value = freq;
        var g = sh.audioContext.createGain();
        g.gain.value = 0.25;
        g.gain.setValueAtTime(0.25, clock.t1);
        g.gain.setTargetAtTime(0.0, clock.t1, decayTime);
        osc.stop(clock.t1 + decayTime * 12);
  • sh.delay(dur, [callback]) will cause the models that follow a delay in a sequence to occur at a later time. If a callback is specified, it will be called like callback(clock, t1r, t2r, startTime, endTime) for every "tick" of the scheduler until the delay finishes. This is useful to perform timed animations.
  • sh.track(model1, model2, ...) makes a sequence for performing the given models one after another. Load the "ping" model above and then run the following code -[ping(440), sh.delay(1.0), ping(660), sh.delay(0.5), ping(880)]));
  • sh.slice(aTrack, startIndex, endIndex) makes a "slice" of the given track that can be played independently.
  • sh.loop(model) makes a model that will loop the given model forever, thus never terminating until a stop is encountered. The following will play a never enging series of "ping"s. (Use sh.cancel() to stop.)[ping(440), sh.delay(0.5)])));
  • sh.loop_while(flag, model) makes a model that will loop as long as the given flag.valueOf() is truthy. Once it becomes falsy, the loop terminates and continues on to the action that follows it in a sequence.
  • sh.repeat(n, model) makes a model that will repeat model the given n number of times. The following code plays 5 pings in sequence -, sh.track([ping(440), sh.delay(0.5)])));
  • sh.fork(model1, model2, ..) causes all the given models to start in "parallel" and waits for them to finish before continuing.
  • sh.spawn(model1, model2, ...) like fork causes all the given models to start in "parallel", but doesn't wait for them to finish.
  • sh.dynamic(function (clock) { return model; }) makes a "dynamic model" that will behave like whatever model the given function returns, at the time the dynamic model gets to run. You can code up temporal recursion using dynamic like this -
    var callback = sh.dynamic(function (clock) {
        // 1. Make some noise at clock.t1
        // 2. Schedule ourselves into the future.
        return sh.track([sh.delay(0.5), callback]);
  • sh.anim(param, duration, v1, v2) animates the parameter from value v1 to value v2 over the given duration using linear interpolation. If param is an AudioParam, then the animation is guaranteed to have sample-accurate timing.
  • sh.anim(param, duration, function (t) { return value; }) will assign the return value of the given function (whose time argument is normalized to [0,1] range) to the parameter over the given duration.
  • sh.anim(param, duration, v1, v2, function (t1) { return t2; }) will use the given interpolation function (domain = [0,1], range = [0,1]) to generate parameter value s between v1 and v2.
  • sh.rate(r) changes the rate of passage of pseudo time through the track in which it occurs to the given r. r = 1 means that the virtual time is itself in seconds.
  • sh.choice(models) makes a model that will randomly behave like one of the models in the given array every time it is invoked. This is a simple use of dynamic.
  • sh.sync() makes a synchronizer that can be used to mark various points within a composition where a particular model's playback can be synced.
    var s1 = sh.sync();
    var p440 = ping(440), p880 = ping(880);
    var halfsec = sh.delay(0.5);
    var trk = sh.track([p440, halfsec, p440, halfsec, s1, halfsec, p440]);; // Start playing the track.; // Will cause model to play when trk gets to s1.
  • sh.gate() makes an action that can be used to pause/resume at certain points within a composition. This is useful for context dependent pause/resume support.
    // Pseudocode
    var g1 = sh.gate();
    var trk = sh.track([model1, model2, ..., g1, ... modelN]);; // Start playing the track.
    g1.close(); // Pause when the play back gets to g1.; // Resume if paused at g1, otherwise equivalent to sh.cont.

The following operators are available for working with visual actions. Audio may be computed a little ahead of time that may span a few visual frames. These functions account for that difference and try to schedule the visual actions as close to the actual render time as possible. These visual actions can be scheduled alongside audio actions in the same composition.

  • sh.display(function (clock, scheduledTime, currentTime) {...}) will call the given callback at the appropriate time for the visual display to occur. The display() action itself takes zero duration just like fire.
  • sh.frame(function (clock) {...}) schedules rendering a single visual frame and lasts as long - i.e. the following action will run only after the frame computation completes. Consecutive frame() actions result in frame by frame synced animation.
  • sh.frames(duration, function (clock) {...}) schedules an animation sequence that lasts for the given duration. It accounts for the fact that any following audio computations may need to be done a little before the visual animation finishes.

If sh.ontick is set to a function (t) {...}, then it will be called for every scheduler tick at the end of all other operations. This is useful for 2D/3D draw flush operations.