sekiguchi92 / soundsourceseparation Goto Github PK

The FastMNMF algorithm uses a lot of memory. A 24s clip at 16kHz with 12 channels (~9.2MB) requires over 25GB of memory.
Can this be dealt with or do the algorithms require the entire memory loaded up-front?

I have some questions.

1. Is there code for FastMNMF2?

2. There doesn't seem to be any requirements on the individual channels in the input file. Do they need to be time synchronised? Is there any restriction on topology of microphones (where there are situated)? Does this topology need to be known a priori?

3. Do the algorithms need to have oversight of the entire data? Or can audio files be batched into blocks, maybe overlapping, to reduce memory usage? Or will that degrade performance/accuracy?

Hello,

Recently I have read your awesome paper "Bayesian Multichannel Speech Enhancement with a Deep Speech Prior" and I want to use your score in the paper. However, when I calculate the average SDR value of Chime-3 development input data, the input SDR is very low which is about -4.5dB, far from 5.8dB reported in the paper. I used BSS_eval python implementation to calculate SDR. Did you use any pre-processing methods to normalize or align the input audio? Or 5.8dB is calculated just from the raw audio?
I would appreciate it if you could give me some suggestions.

What are the maximum number of channels allowed for each algorithm?

@sekiguchi92 Thank you for the latest updates. Do you have plans to build an API, one which can consume numpy arrays instead of filepaths to audio files?