Tag Archives: synthesis

Fragment Synthesizer : GLSL powered HTML5 spectral synthesizer

Some years ago i found out the Virtual ANS synthesizer, this is a very good spectral synthesizer simulating the Russian ANS synthesizer, an unique photoelectronic musical instrument where the score is a drawn sound spectrogram, the x axis of the score represent time and the y axis represent frequency.

Since then i have a great interest in this sort of synthesis and method of composing and have an ongoing large project heavily related to the Virtual ANS.

Fragment Synthesizer is a fun side experiment made quickly where the initial thought was : What if you set GLSL produced images as the source of a spectral synthesizer?

The result is the Fragment Synthesizer web application

This is a full blown stereophonic (the color matter, red for left, green for right) spectral synthesizer which is constantly playing a slice of a GLSL produced image/animation, you can compose by editing the fragment shader or by just copy-pasting code from Shadertoy and then convert it to the Fragment Synthesizer format by clicking on the convert button.

The web app. consist of 3 parts :

  • The score produced by a fragment shader with a vertical bar representing the slice which will be played by the synthesizer
  • A live code editor with the ability to compile as you type (with errors reporting), this is used to edit the GLSL fragment shader and subsequently to compose. (the code editor is powered by the CodeMirror library)
  • Controls (volume slider, button to convert Shadertoy code to Fragment Synthesizer code and a slider to move the playing slice) powered by my own JavaScript widget library

How does it work?

Audio side :

The Fragment Synthesizer is just a simple additive synthesizer under the hood, it is powered by a simple wavetable which is generated with this code :

        _wavetable_size = 32768,
        
        _wavetable = (function (wsize) {
                var wavetable = new Float32Array(wsize),

                    wave_phase = 0,
                    wave_phase_step = 2 * Math.PI / wsize,

                    s = 0;

                for (s = 0; s < wsize; s += 1) {
                    wavetable[s] = Math.sin(wave_phase);

                    wave_phase += wave_phase_step;
                }

                return wavetable;
            })(_wavetable_size),

There is then an oscillator for each lines of the score, oscillators are generated by this function :

    var _generateOscillatorSet = function (n, base_frequency, octaves) {
        var y = 0,
            frequency = 0.0,
            octave_length = n / octaves;
        
        _oscillators = [];

        for (y = n; y >= 0; y -= 1) {
            frequency = base_frequency * Math.pow(2, y / octave_length);

            var osc = {
                freq: frequency,
                
                phase_index: Math.random() * _wavetable_size, 
                phase_step: frequency / _audio_context.sampleRate * _wavetable_size
            };
            
            _oscillators.push(osc);
        }
    };

On the Fragment Synthesizer the starting frequency is 16.34 hertz (bottom of the score) and the y axis span 10 octaves, this is hardcoded but could be fun to let the user change it, the number of oscillators change as the score height change and it depend of the window height, if the user resize the window, the number of oscillators will change.

One of the most important function is the _computeNoteBuffer function, it will transform the pixels array of a vertical slice into an usable and fast to process data structure consisting of a single float32 typed array, each entries (an entry is in fact 5 entries because data is packed linearly into the array) of this array describe which oscillator to play along with data related to how it should play, here is what the 5 entries are :

  • The index of the oscillator to play
  • The previous left side gain value for this oscillator
  • The previous right side gain value for this oscillator
  • The current left side gain value for this oscillator
  • The current right side gain value for this oscillator

The gain value for each side is determined by the red and green component of the pixel value.

The previous gain value is used because it is interpolated when played, this produce a better sound without crackles when the gain value vary greatly.

This function is actually called in the audio callback (it is very fast so ok), it is called for each frames, here is how the number of samples before the next note is computed :

        _fps = 60,
        _note_time = 1 / _fps,
        _note_time_samples = Math.round(_note_time * _sample_rate),

Here is the code of the _computeNoteBuffer function :

    var _computeNoteBuffer = function () {
        for (i = 0; i < _note_buffer.length; i += 1) {
            _note_buffer[i] = 0;
        }
        
        var note_buffer = _note_buffer,
            pvl = 0, pvr = 0, pr, pg, r, g,
            inv_full_brightness = 1 / 255.0,

            dlen = _data.length,
            y = _canvas_height - 1, i,
            volume_l, volume_r,
            index = 0;

        for (i = 0; i < dlen; i += 4) { pr = _prev_data[i]; pg = _prev_data[i + 1]; r = _data[i]; g = _data[i + 1]; if (r > 0 || g > 0) {
                volume_l = r * inv_full_brightness;
                volume_r = g * inv_full_brightness;
                
                pvl = pr * inv_full_brightness;
                pvr = pg * inv_full_brightness;

                note_buffer[index] = y;
                note_buffer[index + 1] = pvl;
                note_buffer[index + 2] = pvr;
                note_buffer[index + 3] = volume_l - pvl;
                note_buffer[index + 4] = volume_r - pvr;
            } else {
                if (pr > 0 || pg > 0) {
                    pvl = pr * inv_full_brightness;
                    pvr = pg * inv_full_brightness;

                    note_buffer[index] = y;
                    note_buffer[index + 1] = pvl;
                    note_buffer[index + 2] = pvr;
                    note_buffer[index + 3] = -pvl;
                    note_buffer[index + 4] = -pvr;
                }
            }

            y -= 1;

            index += 5;
        }
        
        _prev_data = _data;
        
        _swap_buffer = true;
    };

Now the core audio code where the magic happen (nothing really fancy here except the interpolation) :

    var _audioProcess = function (audio_processing_event) {
        var output_buffer = audio_processing_event.outputBuffer,
            
            output_data_l = output_buffer.getChannelData(0),
            output_data_r = output_buffer.getChannelData(1),
            
            output_l = 0, output_r = 0,
            
            wavetable = _wavetable,
            
            note_buffer = _note_buffer,
            note_buffer_len = note_buffer.length,
            
            wavetable_size_m1 = _wavetable_size - 1,
            
            osc,
            
            lerp_t_step = 1 / _note_time_samples,
            
            sample,
            
            s, j;
        
        for (sample = 0; sample < output_data_l.length; sample += 1) {
            output_l = 0.0;
            output_r = 0.0;

            for (j = 0; j < note_buffer_len; j += 5) { var osc_index = note_buffer[j], previous_volume_l = note_buffer[j + 1], previous_volume_r = note_buffer[j + 2], diff_volume_l = note_buffer[j + 3], diff_volume_r = note_buffer[j + 4]; osc = _oscillators[osc_index]; s = wavetable[osc.phase_index & wavetable_size_m1]; output_l += (previous_volume_l + diff_volume_l * _lerp_t) * s; output_r += (previous_volume_r + diff_volume_r * _lerp_t) * s; osc.phase_index += osc.phase_step; if (osc.phase_index >= _wavetable_size) {
                    osc.phase_index -= _wavetable_size;
                }
            }
            
            output_data_l[sample] = output_l;
            output_data_r[sample] = output_r;
            
            _lerp_t += lerp_t_step;
            
            _curr_sample += 1;

            if (_curr_sample >= _note_time_samples) {
                _lerp_t = 0;

                _curr_sample = 0;

                _computeNoteBuffer();
            }
        }
    };
Visual side :

On the visual side, it is powered by WebGL and GLSL, there is a basic screen aligned quad with a fragment shader applied to it and then for each frames a readPixels call is made to get the user chosen (_play_position) vertical slice pixels array which will get converted in the audio callback by the computeNoteBuffer function, here is the code called for each frames :

    var _frame = function (raf_time) { 
        _gl.useProgram(_program);
        _gl.uniform1f(_gl.getUniformLocation(_program, "globalTime"), (raf_time - _time) / 1000);
        _gl.uniform2f(_gl.getUniformLocation(_program, "iMouse"), _mx, _my);

        _gl.drawArrays(_gl.TRIANGLE_STRIP, 0, 4);

        if (_swap_buffer) {
            _gl.readPixels((_canvas_width - 1) * _play_position, 0, 1, _canvas_height, _gl.RGBA, _gl.UNSIGNED_BYTE, _data);

            _swap_buffer = false;
        }

        _raf = window.requestAnimationFrame(_frame);
    };

And voilà, the core of the synthesizer explained.

Note : Borrowed from Shadertoy, the resolution, iMouse and globalTime uniforms are defined and can be used in the fragment shader to do the cool stuff! 🙂

The full source code is available on github

1 Star2 Stars3 Stars4 Stars5 Stars (No Ratings Yet)
Loading...