Imported from matlab/matlab-agentic-toolkit (skills-catalog/signal-processing/matlab-process-streaming-audio/SKILL.md). Install upstream with npx skills add matlab/matlab-agentic-toolkit --skill matlab-process-streaming-audio. Copyright stays with the author (https://www.mathworks.com/content/dam/ma).

Streaming Audio Processing

Design and run real-time audio processing in MATLAB and Simulink using Audio Toolbox streaming objects. These objects maintain internal state across frames, support tunable properties, and provide built-in visualization.

When to Use

• Building frame-based audio processing loops
• Filtering audio in real time (crossover, EQ, shelving, octave)
• Applying dynamic range control (compressor, limiter, expander, noise gate)
• Measuring audio levels (peak, loudness, SPL, octave-band spectra)
• Resampling audio signals (sample rate conversion)
• Applying long impulse responses in real time (frequency-domain filtering)
• Tuning audio parameters interactively while streaming
• Visualizing filter responses or compressor characteristics
• Controlling audio parameters with MIDI devices
• Building audio processing chains in Simulink

When NOT to Use

• Audio device I/O setup with audiostreamer — use the matlab-play-record-audio skill
• Audio plugin generation (VST/AU via createAudioPluginClass)
• Deep learning audio features or inference
• Offline batch processing of entire files without streaming

Workflow

Every streaming audio task follows this pattern:

1. Create source — dsp.AudioFileReader (or audiostreamer for live I/O)
2. Create processing objects — Audio Toolbox System objects configured for your sample rate
3. Visualize responses — call visualize(obj) on filter/DRC objects
4. Open tuning UI — call parameterTuner(obj) for interactive control
5. Process in a loop — read frames, process, write output
6. Clean up — release all objects

% Standard streaming audio processing pattern
reader = dsp.AudioFileReader("input.wav", SamplesPerFrame=256);
fs = reader.SampleRate;

crossFilt = crossoverFilter(2, [500 4000], SampleRate=fs);
comp = compressor(Threshold=-20, Ratio=4, SampleRate=fs);

visualize(crossFilt);
visualize(comp);
parameterTuner(crossFilt);
parameterTuner(comp);

while ~isDone(reader)
    audioIn = reader();
    [low, mid, high] = crossFilt(audioIn);
    low = comp(low);
    audioOut = low + mid + high;
    drawnow limitrate  % flush UI events so parameterTuner changes take effect
end

release(reader);
release(crossFilt);
release(comp);

Key Functions

Use These (Audio Toolbox objects)

Universal Methods

MIDI Control Functions

Patterns

Visualization and Tuning

parameterTuner works on every Audio Toolbox streaming object. visualize works on most objects — exceptions: reverberator and splMeter (use parameterTuner instead). Always use these methods instead of building custom UIs.

eq = multibandParametricEQ(NumEQBands=5, ...
    HasLowShelfFilter=true, HasHighShelfFilter=true, SampleRate=fs);

% One-line visualization — shows combined magnitude response
visualize(eq);

% One-line interactive tuning — sliders for all tunable properties
parameterTuner(eq);

Objects supporting parameterTuner: compressor, expander, limiter, noiseGate, octaveFilter, crossoverFilter, multibandParametricEQ, graphicEQ, audioOscillator, wavetableSynthesizer, reverberator, shelvingFilter, octaveSpectrumEstimator.

Multiband Processing

Split → process per band → sum. Use crossoverFilter for the split.

crossFilt = crossoverFilter(2, [500 4000], 48, fs);  % 2 crossovers, 48 dB/oct
compLow = compressor(Threshold=-20, Ratio=4, SampleRate=fs);
compMid = compressor(Threshold=-15, Ratio=3, SampleRate=fs);
compHigh = compressor(Threshold=-10, Ratio=2, SampleRate=fs);

% In the processing loop:
[low, mid, high] = crossFilt(audioIn);
audioOut = compLow(low) + compMid(mid) + compHigh(high);

Parametric Equalization

Use multibandParametricEQ for streaming EQ. It supports N bands, optional low/high shelves, optional lowpass/highpass, and oversampling.

eq = multibandParametricEQ( ...
    NumEQBands=5, ...
    EQOrder=4, ...
    Frequencies=[100 400 1000 4000 8000], ...
    QualityFactors=[0.7 1.5 2.0 1.8 0.7], ...
    PeakGains=[3 -2 4 -1.5 2], ...
    HasLowShelfFilter=true, LowShelfCutoff=80, LowShelfGain=2, ...
    HasHighShelfFilter=true, HighShelfCutoff=12000, HighShelfGain=-1, ...
    SampleRate=fs);

visualize(eq);
parameterTuner(eq);

% In the loop — all properties are tunable while streaming:
audioOut = eq(audioIn);

For coefficient-level control (e.g., feeding a dsp.SOSFilter), use design functions:

% designParamEQ — name-value syntax
[B, A] = designParamEQ(CenterFrequency=1000/(fs/2), ...
    QualityFactor=2, Gain=6, FilterOrder=4);

% designShelvingEQ — positional syntax: (gain, slope, normalizedFc, type)
[B, A] = designShelvingEQ(3, 0.8, 200/(fs/2), "lo", Orientation="row");

% designVarSlopeFilter — positional syntax: (slope, normalizedFc, type)
[B, A] = designVarSlopeFilter(24, 5000/(fs/2), "lo", Orientation="row");

Metering

Choose the metering object based on what you are measuring:

% Digital peak level meter (most common "give me a level meter" answer)
lvl = audioLevelMeter(Method="true-peak", SampleRate=fs);
visualize(lvl);
while ~isDone(reader)
    lvl(reader());
    drawnow limitrate
end

% Broadcast loudness meter (EBU R128)
loud = loudnessMeter(SampleRate=fs);
visualize(loud);
while ~isDone(reader)
    loud(reader());
    drawnow limitrate
end

% octaveSpectrumEstimator — preferred for octave-band visualization
ose = octaveSpectrumEstimator(fs, ...
    Bandwidth="1/3 octave", ...
    FrequencyWeighting="A-weighting", ...
    TimeWeighting="fast");

visualize(ose);         % Built-in real-time bar chart
parameterTuner(ose);    % Tune bandwidth, weighting, etc. while streaming

while ~isDone(reader)
    audioIn = reader();
    [spectrum, centerFreqs] = ose(audioIn);
end

% splMeter — when you need Lt, Leq, Lpeak, Lmax outputs
% NOTE: Initial frames return -Inf until the time-weighted filter
% accumulates sufficient energy (~10-50 frames). This is normal.
spl = splMeter( ...
    Bandwidth="1/3 octave", ...
    FrequencyWeighting="A-weighting", ...
    TimeWeighting="fast", ...
    SampleRate=fs);

[Lt, Leq, Lpeak, Lmax] = spl(audioIn);

Streaming Spectral Processing (Per-Bin Manipulation)

Use dsp.STFT + dsp.ISTFT when you need to manipulate individual frequency bins in a streaming loop (spectral gating, spectral subtraction, phase vocoder effects). These objects handle windowing, overlap, buffering, and perfect reconstruction internally.

% Streaming spectral noise gate using dsp.STFT / dsp.ISTFT
fftLen = 1024;
overlapLen = fftLen * 3/4;  % 75% overlap
win = hann(fftLen, 'periodic');

stf = dsp.STFT(win, overlapLen, fftLen);
istf = dsp.ISTFT(win, overlapLen);

reader = dsp.AudioFileReader("input.wav", SamplesPerFrame=fftLen-overlapLen);
writer = dsp.AudioFileWriter("output.wav", SampleRate=reader.SampleRate);

while ~isDone(reader)
    audioIn = reader();
    X = stf(audioIn);           % Windowed FFT with overlap handled
    X(abs(X) < threshold) = 0;  % Per-bin manipulation
    audioOut = istf(X);          % Perfect-reconstruction IFFT + OLA
    writer(audioOut);
end

release(reader); release(writer); release(stf); release(istf);

Ranked by idiom quality for spectral tasks:

1. dsp.STFT + dsp.ISTFT — handles all buffering, windowing, COLA internally
2. dsp.AsyncBuffer for manual buffering + fft/ifft
3. Fully hand-rolled buffer shifting (avoid — error-prone, no COLA guarantee)

Important: dsp.FrequencyDomainFIRFilter does NOT expose per-bin access. It applies a fixed FIR (impulse response) in the frequency domain. Use it for convolution reverb and long IR filtering, not for spectral manipulation.

Frequency-Domain FIR Filtering (Long Impulse Responses)

Use dsp.FrequencyDomainFIRFilter when you need to convolve with a long, fixed impulse response in a streaming loop — room IRs, cabinet IRs, or any scenario where time-domain convolution would be too slow for real-time. The object implements overlap-add (or overlap-save) internally and maintains state across frames.

% Stream audio through a long impulse response (e.g., room IR)
[ir, irFs] = audioread("impulse_response.wav");
ir = ir.';  % Numerator must be a row vector
reader = dsp.AudioFileReader("input.wav", SamplesPerFrame=1024);
fs = reader.SampleRate;
writer = audioDeviceWriter(SampleRate=fs);

fdFilt = dsp.FrequencyDomainFIRFilter(ir, ...
    PartitionForReducedLatency=true, ...
    PartitionLength=1024);

while ~isDone(reader)
    audioIn = reader();
    audioOut = fdFilt(audioIn);
    writer(audioOut);
end

release(reader);
release(fdFilt);
release(writer);

When to use which frequency-domain approach:

Partitioned convolution: Set PartitionForReducedLatency=true and PartitionLength to your frame size (e.g., 1024). This splits the IR into partitions, reducing latency to one partition instead of the full IR length — critical for real-time applications.

Real-Time Pacing for File-Based Loops

When reading from a file with visualizations or metering, the loop runs at full CPU speed — frames fly past faster than the display can render. Add pacing so the visualization is meaningful:

frameDuration = reader.SamplesPerFrame / fs;
while ~isDone(reader)
    audioIn = reader();
    % ... process ...
    loud(audioOut);
    drawnow limitrate
    pause(frameDuration);  % Pace to approximately real time
end

Alternatively, use audioDeviceWriter which inherently blocks to maintain real-time pacing (audio plays through speakers at the correct rate). If you only need file output without real-time playback, pacing is unnecessary.

Vectorized Per-Frame Gain Smoothing

When applying exponential smoothing toward a constant target within a frame, avoid per-sample for loops. The gain trajectory is a geometric series with a closed-form solution:

% Instead of per-sample loop:
n = (1:numSamples)';
gainVector = targetGain + (currentGain - targetGain) * smoothingCoeff.^n;
audioOut = audioIn .* gainVector;
currentGain = gainVector(end);

This is significantly faster than iterating sample-by-sample and produces identical results when the target gain is constant across the frame.

Sample Rate Conversion

Use audioresample (R2023b) for one-shot conversion or designAudioResampler for streaming.

% One-shot (entire signal)
audioOut = audioresample(audioIn, InputRate=96000, OutputRate=44100);

% Streaming — design once, use in loop
resampler = designAudioResampler(InputRate=44100, OutputRate=16000);
% resampler is a dsp.FIRRateConverter or dsp.FilterCascade — use in loop:
audioOut = resampler(audioFrame);

SamplesPerFrame alignment: When streaming with designAudioResampler, set SamplesPerFrame on the reader to a multiple of the resampler's DecimationFactor. This ensures every output frame has a consistent, fixed length. If the frame size is not aligned, output frames vary in length, which breaks downstream fixed-frame processing.

% Example: 44100 Hz → 16000 Hz
resampler = designAudioResampler(InputRate=44100, OutputRate=16000);
% DecimationFactor is 441 — set SamplesPerFrame to a multiple of 441
reader = dsp.AudioFileReader("input.wav", SamplesPerFrame=441);

MIDI Control

Use MIDI devices to tune parameters in real time during streaming.

% Quick approach — midicontrols for reading control values
controls = midicontrols(1:3);  % 3 MIDI controls (auto-detect with midiid)
while ~isDone(reader)
    vals = midiread(controls);  % Returns values in [0, 1]
    comp.Threshold = -60 + vals(1) * 60;   % Map to [-60, 0] dB
    comp.Ratio = 1 + vals(2) * 19;         % Map to [1, 20]
    audioOut = comp(reader());
end

% Full approach — mididevice for send/receive
device = mididevice("Oxygen 49");
msgs = midireceive(device);
midisend(device, midimsg("ControlChange", 1, 64, 100));

Simulink Audio Chain

See references/simulink-audio-blocks.md for the full block catalog. Key setup:

% Solver: fixed-step discrete, auto step size
set_param(model, 'Solver', 'FixedStepDiscrete');
set_param(model, 'FixedStep', 'auto');

% Blocks are in these libraries:
%   audiosources  — From Multimedia File, Audio Device Reader, MIDI Controls
%   audiosinks    — Audio Device Writer, Spectrum Analyzer, To Multimedia File
%   audiofilters  — Crossover Filter, Multiband Parametric EQ, Graphic EQ,
%                   Octave Filter, Shelving Filter, Weighting Filter,
%                   Parametric EQ Design, Shelving EQ Design, Variable Slope Filter Design
%   audiodynamicrange — Compressor, Expander, Limiter, Noise Gate
%   audioeffects  — Reverberator

Design blocks (Parametric EQ Design, Shelving EQ Design, Variable Slope Filter Design) separate filter design from implementation — they output coefficients to feed a SOS/FOS filter block. Use them when:

• You need codegen-friendly architectures
• You want to specify bandwidth by octave or band-edge frequencies
• You need higher-order filters than the integrated blocks support

"Visualize Response" button — filter and DRC blocks have a built-in button on their dialog that shows the response while the model runs and updates as parameters change. No extra blocks needed.

Spectrum Analyzer from audiosinks — use this instead of the DSP library version. For audio-friendly settings, configure: one-sided spectrum, log frequency scale.

Common Mistakes

Conventions

• Always set SampleRate explicitly on Audio Toolbox objects — never rely on the 44100 default
• Always include drawnow limitrate inside the processing loop when using parameterTuner or any interactive UI — without it, MATLAB never processes UI events and tuning changes do not take effect
• Set PlayCount=Inf on dsp.AudioFileReader when the user needs time to interact with tuning UIs or visualizations — the default PlayCount=1 terminates too quickly for interactive use
• Use timescope and spectrumAnalyzer for real-time signal viewing (one-sided spectrum, log frequency axis for audio)
• Prefer crossoverFilter slopes of 24 or 48 dB/octave (Linkwitz-Riley alignment)
• Use isDone(reader) to control streaming loops, not manual frame counting
• Call release(obj) on all objects after the loop completes
• For Simulink: always use FixedStepDiscrete solver with FixedStep='auto'

Copyright 2026 The MathWorks, Inc.