A python implementation of the Griffin Lim Algorithm for audio reconstruction from magnitudes
The Griffin Lim algorithm provides a way to approximate a waveform from a "modified STFT transform". In practice, this means a STFT where the phase data has been lost or removed. The Griffin Lim Algorithm does not produce the original waveform, but rather a new waveform with a very similar spectrogram.
Improvements on the Modified Fast Griffin Lim are still possible, as evidenced by recent work with WaveNet. However, in practice, the Modified Griffin Lim remains a robust, efficient and understandable algorithm, which makes it useful in many audio pipelines.
This is an implementation of 3 different versions of the Griffin Lim algorithm.
- The basic Griffin Lim which converges slowly
- The Fast Griffin Lim which converges more quickly
- The Modified Fast Griffin Lim which converges to slightly better values.
After installing this package with pip, you can use the command pygl which has reasonable help.
I have not read the Griffin Lim paper very closely, and I expect that my implementation is missing crucial aspects. But this is a practical repository, not an academic one. Please report any bugs as github issues.