jurihock / qdft Goto Github PK

Variable-Q transform is very similar to constant-Q transform except the Q value is lower as the frequency decreases, which is useful if you want to have better time resolution on lower frequencies at the expense of frequency resolution at bass frequencies (like a logarithmic-frequency spectrogram with ERB frequency resolution).

Arbitrary frequency scale for bin spacing also enables variable-Q transform and with that, it can directly calculate Mel spectrogram using VQT with Mel-frequency bin spacing and resolution.

It is worth noting that asymmetric window functions are impossible to implement, not even the window skew parameter outlined in my collection of audio analysis functions on a sliding DFT, but what you can do is make the sDFT an IIR filter bank (which is a damped sine wave). which uses less RAM (since it doesn't need ring buffers) and it can be cascaded (just like biquad filters) to achieve a steeper rolloff at the cost of greater delay (thus, bad time resolution) on long window sizes

• Deep Salience Representations for f0 Estimation in Polyphonic Music
• A Lightweight Instrument-Agnostic Model for Polyphonic Note Transcription and Multipitch Estimation
• pumpp.feature.HCQT

As mentioned in my comment, the std::rotate is most likely a major performance issue.

Actually there is no measurable difference between the following variations:

• std::rotate(inputs.begin(), inputs.begin() + 1, inputs.end());
• std::copy(inputs.begin() + 1, inputs.end(), inputs.begin());
• std::shift_left(inputs.begin(), inputs.end(), 1);

But instead of std::vector use std::deque to shift the buffered values like so:

• inputs.pop_front();
• inputs.push_back(sample);

The complexity of the linear/random element access as well as insertion or removal of elements at both ends of the std::deque should be still O(1).

Check this one:

• https://ieeexplore.ieee.org/document/9770278
• https://github.com/zafarrafii/CQHC-Python

See also:

• https://en.wikipedia.org/wiki/Harmonic_series_(music)#Harmonics_and_tuning

• Basic usage as a standalone example
• Additional getters
• C++ namespace qdft
• Rework convolution loop
• Try to build all C++ examples as possible, regardless of the Python availability
• Bin latency values in seconds

as implemented in jurihock/sdft

(see also jurihock/sdft#5)

see also Fast, Accurate Frequency Estimators

Just do it and publish on crates.io