AudioEffectPhaseVocoder

A phase vocoder audio effect for Teensy 4.x, forked from duff2013/AudioEffectVocoder by Colin Duffy.

About this fork

This fork adds time-stretching, pitch shifting, and reverse playback — playing audio at variable speed and pitch without the constraints of standard sample playback. The effect owns its sample buffer directly (loaded from SD or PSRAM) and integrates with the Teensy Audio Library as a 0-input / 1-output node.

Target hardware: Teensy 4.0 / 4.1 + Audio Shield (SGTL5000). Requires Teensy 4 due to the CMSIS 5 FFT API (arm_cfft_f32).

API

AudioEffectPhaseVocoder vocoder;

// Load a sample (mono 16-bit PCM, 44100 Hz)
vocoder.setSample(sampleBuffer, numSamples);

// Playback control
vocoder.play();
vocoder.stop();
vocoder.setLoop(true);
bool playing = vocoder.isPlaying();

// Time-stretching
// > 1.0 = slower (e.g. 2.0 = double duration)
// < 1.0 = faster (e.g. 0.5 = half duration)
vocoder.setStretch(1.5f);
float s = vocoder.getStretch();

// Pitch shifting (semitones). 0 = no shift, +12 = octave up, -12 = octave down.
// Independent of time-stretching — both can be applied simultaneously.
vocoder.setPitchShift(3.0f);
float p = vocoder.getPitchShift();

// Reverse playback — plays the sample backwards through the phase vocoder.
// Time-stretching and pitch shifting apply in reverse as normal.
vocoder.setReverse(true);
bool r = vocoder.isReverse();

// Transient detection sensitivity (default 16.0)
// Lower = more phase resets on attacks → better transient preservation
// Higher = fewer resets → smoother sustain
vocoder.setTransientThreshold(16.0f);

// Performance profiling — returns mean and peak update() time in µs since last call.
float meanUs; uint32_t peakUs;
if (vocoder.getProfiling(meanUs, peakUs)) { ... }

Examples

timestretch_example.ino

Single-voice sketch. Loads up to three WAV files from SD card and plays them with real-time time-stretch and pitch control over Serial.

Key	Action
SPACE	Start / restart
s	Stop
1–3	Load 01.WAV / 02.WAV / 03.WAV
p	Faster
q	Slower
w	Pitch up 1 semitone
x	Pitch down 1 semitone
z	Reset pitch to 0
r	Toggle reverse
t	Transient threshold: 4 (sensitive)
y	Transient threshold: 8
u	Transient threshold: 16 (subtle)
d	Toggle profiling report
h	Help

poly_example.ino

Four-voice polyphonic sketch. Loads 10 WAV files into PSRAM at startup. Each voice is independently triggerable with per-sample persistent settings. Uses steal-oldest voice allocation with AudioEffectFade crossfades.

Key	Action
1–9, 0	Trigger sample 1–10
SPACE	Re-trigger last selected sample
s	Stop all voices
p	Faster (last selected sample)
q	Slower (last selected sample)
w	Pitch up 1 semitone
x	Pitch down 1 semitone
z	Reset pitch to 0
r	Toggle reverse
l	Toggle loop
t/y/u	Transient threshold 4 / 8 / 16 (all voices)
d	Toggle profiling report (2 s interval)
h	Help

Controls apply to the most recently triggered sample. Settings persist between triggers.

3 test WAV files are included in the test/ directory — copy them to an SD card.

Memory

The DSP arrays (~67 KB) live in fast DTCM (RAM1). Sample buffers should be placed in OCRAM or PSRAM to avoid overflowing DTCM:

DMAMEM int16_t sampleBuffer[44100 * 5];   // Teensy 4.0/4.1 — OCRAM (~430 KB)
EXTMEM int16_t sampleBuffer[44100 * 60];  // Teensy 4.1 + PSRAM — up to ~8 MB

How it works

The phase vocoder uses a Short-Time Fourier Transform (STFT) to analyse overlapping frames of audio, manipulate the phase spectrum, and reconstruct the output via overlap-add.

Time-stretching decouples the analysis hop (how fast input is consumed) from the synthesis hop (always 128 samples, locked to Teensy's audio block rate):

stretch_factor = synthesis_hop / analysis_hop

• stretch > 1 → smaller analysis hop → reads input slowly → longer output
• stretch < 1 → larger analysis hop → reads input fast → shorter output

Reverse playback rebuilds the analysis window each frame by reading samples in descending order (window[j] = sample[read_pos - j]), giving the FFT a genuinely time-reversed view. Phase advances in the same direction as forward playback, so all existing phase accumulation logic applies unchanged. At unity stretch and pitch, reverse falls through to a lightweight passthrough path.

Transient detection via spectral flux compares the instantaneous sum of positive magnitude increases against a smoothed reference. On a detected transient, Phase_Sum is set to the current analysis phase rather than accumulated, which preserves harmonic coherence on drum hits and attacks.

FFT size: 1024 — Overlap: 8× — Platform: Teensy 4.0 / 4.1

Credits

• Original vocoder: Colin Duffy
• Phase vocoder algorithm based on Stephan M. Bernsee's smbPitchShift