ExtremaSpectrum 1.0.0

ExtremaSpectrum

ExtremaSpectrum is a .NET library for audio signal analysis based on an extrema-driven oscillation decomposition algorithm.

This is not an FFT spectrum. The output describes detected local oscillations by frequency, not orthogonal sinusoidal components. See Algorithm.

Repository: github.com/DimonSmart/ExtremaSpectrum

Visual overview

Visual frequency inspection for sound.
ExtremaSpectrum decomposes a waveform into local oscillations, making dominant frequency bands and each reduction pass easy to inspect.

Try the demo with a low-frequency focused view:

dotnet run --project src\ExtremaSpectrum.Demo -- --min-frequency 0 --input ".\data\demo-low-010hz-plus-high-160hz.wav" --bins 180 --to-bin 20

AnalyzeDetailed(...) returns an ExtremaAnalysisReport with per-pass spectra, accepted oscillations, support ranges, and waveform snapshots.
The demo app can also export step-by-step SVG frames:

dotnet run --project src/ExtremaSpectrum.Demo -- --input data/demo-low-010hz-plus-high-160hz.wav --export-step-images temp

Example frames generated from data/demo-low-010hz-plus-high-160hz.wav:

Install

dotnet add package ExtremaSpectrum

The current package targets .NET 10.

Quick start

using ExtremaSpectrum;

var analyzer = new ExtremaSpectrumAnalyzer(new ExtremaSpectrumOptions
{
    BinCount = 128,
    MinFrequencyHz = 500f,
    MaxFrequencyHz = 8000f,
    MaxPasses = 12
});

AnalysisResult result = analyzer.Analyze(samples, sampleRate);
float[] spectrum = result.Spectrum;

Streaming

var analyzer = new StreamingExtremaSpectrumAnalyzer(
    options,
    analysisWindowSamples: 2048,
    hopSamples: 512);

if (analyzer.PushPcm16(buffer, format, out var result))
{
    float[] spectrum = result!.Spectrum;
}

if (analyzer.PushDetailedPcm16(buffer, format, out var report))
{
    IReadOnlyList<ExtremaPassSnapshot> passes = report!.Passes;
}

Algorithm

The analyzer performs iterative decomposition of a discrete signal into local oscillations.

One pass

1. Find all local extrema using strict neighbour comparison. Boundary points are excluded.
2. Scan left to right for consecutive triples in the form min -> max -> min or max -> min -> max.
3. For each accepted triple (left, mid, right):
   • periodSamples = right - left
   • L = floor(midpoint(left, mid))
   • R = ceil(midpoint(mid, right))
   • when the next accepted triple reuses the same slope, its left boundary starts from the already chosen R
   • baseline = (signal[left] + signal[right]) / 2
   • amplitude = abs(signal[mid] - baseline)
   • frequencyHz = sampleRate / periodSamples
4. If the oscillation passes the period and amplitude filters and maps into a valid bin, its contribution is added to the spectrum.
5. Samples strictly between L and R are removed from later passes, while the midpoint boundary samples are preserved.
6. The scan advances by one extremum, so adjacent accepted oscillations can share the same preserved midpoint boundary.

Multiple passes are performed until no valid triple is found or MaxPasses is reached.

Extrema detection

local max: signal[i - 1] < signal[i] && signal[i] >= signal[i + 1]
local min: signal[i - 1] > signal[i] && signal[i] <= signal[i + 1]

Flat plateaus use the first point that satisfies the comparison.

Frequency binning

binWidth = (MaxFrequencyHz - MinFrequencyHz) / BinCount
binIndex = floor((frequencyHz - MinFrequencyHz) / binWidth)

Input formats

All analysis overloads need audio samples plus a sample rate in Hz. The difference is how decoded the input already is.

AudioBufferFormat describes how byte buffers are interpreted:

var format = new AudioBufferFormat
{
    SampleRate = 48000,
    Channels = 2,
    BitsPerSample = 16,
    Interleaved = true,
    ChannelMixMode = ChannelMixMode.AverageAllChannels
};

AnalysisResult result = analyzer.AnalyzePcm16(buffer, format);

Key AudioBufferFormat fields:

• SampleRate: samples per second, for example 44100 or 48000
• Channels: total channel count, for example 1 for mono or 2 for stereo
• BitsPerSample: 16 for AnalyzePcm16(...), 32 for AnalyzeFloat32(...)
• Interleaved = true: samples are laid out as L0 R0 L1 R1 ...
• Interleaved = false: planar layout, for example all left samples followed by all right samples
• PreferredChannel: zero-based channel index used only when ChannelMixMode is PreferredChannel

AnalyzeDetailed(...) follows the same input rules as Analyze(...). The streaming analyzer uses the same PCM buffer rules for PushPcm16(...) and PushFloat32(...).

Multi-channel buffers are mixed to mono according to ChannelMixMode:

Configuration

new ExtremaSpectrumOptions
{
    BinCount = 128,
    MinFrequencyHz = 100f,
    MaxFrequencyHz = 8000f,
    MaxPasses = 16,
    MinPeriodSamples = 2,
    MaxPeriodSamples = 0,
    MinAmplitude = 0f,
    AccumulationMode = AccumulationMode.Amplitude
}

Limitations

• Not FFT. Bin values are accumulated oscillation amplitudes or energies, not DFT coefficients.
• There is no windowing, zero-padding, or spectral leakage correction.
• Frequency resolution improves with longer input buffers.
• The greedy left-to-right pass does not resolve overlap inside the same pass.
• Results depend on signal amplitude. Normalize input if absolute comparison matters.

License

MIT