dsp::Oversampling magnitude

General JUCE discussion
1

I noticed that the magnitude of the signal is increased in the oversampled buffer.
is that a bug or am I missing something?

here attached is a PIP plugin with the following process code that asserts when checking that the magnitude in the oversampled block is <= 1

OversamplingTest.h (6,2 Ko)

TestProcessor()
: AudioProcessor (BusesProperties().withInput  ("Input",  AudioChannelSet::stereo())
                  .withOutput ("Output", AudioChannelSet::stereo()))
{
    oversampling = new dsp::Oversampling<float> (2, 1,
                                                 dsp::Oversampling<float>::filterHalfBandFIREquiripple,
                                                 true);
}

//==============================================================================
void prepareToPlay (double, int samplesPerBlock) override
{
    oversampling->initProcessing ((size_t) samplesPerBlock);
    oversampling->reset();
}

void processBlock (AudioBuffer<float>& buffer, MidiBuffer&) override
{
    int numChannels = buffer.getNumChannels();
    int numSamples = buffer.getNumSamples();

    // overwrite the buffer with noise
    for (int i = 0; i < numChannels; ++i)
    {
        auto* buf = buffer.getWritePointer (i);

        for (int j = 0; j < numSamples; ++j)
            buf[j] = random.nextFloat() * 2.f - 1.f;
    }

    // check the magnitude in our buffer
    {
        auto r = FloatVectorOperations::findMinAndMax (buffer.getReadPointer (0),
                                                       (int) buffer.getNumSamples());
        auto magnitude = jmax (r.getStart(), -r.getStart(), r.getEnd(), -r.getEnd());
        jassert (magnitude <= 1.f);
    }

    // Upsampling
    dsp::AudioBlock<float> block (buffer);
    dsp::AudioBlock<float> oversampledBlock = oversampling->processSamplesUp (block);

    // check the magnitude in the oversampled block
    {
        auto r = FloatVectorOperations::findMinAndMax (oversampledBlock.getChannelPointer (0),
                                                       (int) oversampledBlock.getNumSamples());
        auto magnitude = jmax (r.getStart(), -r.getStart(), r.getEnd(), -r.getEnd());
        jassert (magnitude <= 1.f); // ASSERTS !
    }

    // Downsampling
    oversampling->processSamplesDown (block);

    // clear buffer
    buffer.clear();
}
2

This code might just detect inter-sample peaks. For a proper check you’d need to calculate rms values or calculate the difference between input and output.

3

This can happen when oversampling. Imagine a sine wave being sampled right before and after its peak. The original magnitude is higher than each of the sample values. With oversampling you basically create a value in-between, which of course should be closer to the original peak.

I haven’t tested your code, but the difference between both magnitudes shouldn’t be large, maybe around 1-3 dB max?

1 Like
4

well, if i just added that :

if (magnitude > peak)
{
    peak = magnitude;
    DBG ("peak " << peak);
}

and got :

peak 1.45538
peak 1.63832
peak 1.70835
peak 1.7943
peak 1.80937
peak 1.82301
peak 1.82725
peak 1.83453
peak 1.84099

so that’s more ~5dB peaks

5

Still okay I’d say. Especially as you are testing it with white noise which has a good amount of high frequencies compared to natural audio/music signals which are kind of pink. The high frequencies cause the higher inter-sample-peaks as their slope is steeper.

Have you compared the peaks also with what you get after downsampling?

6

yes, that’s right.
after downsampling the peak gain is ok of course in my example.
But I was concerned about what can happen while processing the oversampled buffer:
if you have a non-linear processing going on, then the result can eventually be really different as the peak gain of the input signal varies.

7

And this difference is exactly why you’d want to do oversampling when doing non-linear dsp. The oversampled version is closer to what would happen with a continous signal. The non oversampled version introduces aliasing

as the peak gain of the input signal varies.

the true peak does not really vary between the two though.

8

thanks for the clarification guys