Convolver DSP plug-in for Windows Media Player 10 - Page 2

Is there any way that you can let me have your filter, please, so that I can try it?

Ed's measuring tool is windows-executable, if you look around the duffroomcorrection site.

The Jones Guide leads you to a PCM file. To get it to work with Convolver, you will need to save it as a WAV file.

I have just put up on http://convolver.sourceforge.net/ a new release (1.19) that offers the option of using partitioned convolution, with reduced lag and possible performance improvement.

Ed's RecImp tool can be compiled into a Windows executable.

Others have tried the DRC route, and tried to combine the individual channel impulse responses that it produces into a single (5.1) WAV file using Audition. Unfortunately the Audition-generated WAV file does not seem to be interpreted correctly by Convolver.

I will investigate this further. In any case, I need to replace the Microsoft-provided example file that I use for reading WAVs with a GPL equivalent, such as libsndfile or libaudiofile.

Feedback welcome.

A further build (1.21) has some more optimizations.

The optimum attenuation is now calculated correctly.

There are also a number of command line utilities (eg, perftest to allow you to see how your filter performs with different partition lengths). and makeIR allows you to generate a filter from a text file specification).

Let me know how you get on with it.

Although Convolver will be visible to ffdshow, I now want to spend more time on the DirectShow wrapper and to recognise a wider range of filter file formats. I have at least one example of multi-channel files generated by Audition not being interpreted correctly.

Happy convolving / duff room correction.

I have just put a further release up (1.23).

It supports a greater variety of filter file formats and is slightly faster. See the notes on http://convolver.sourceforge.net

I suggest that you try this release (or 1.21) as 1.22 was broken when used with more than one partition.

Creating stereo or multi-channel wav files from mono files can be accomplished with Adobe Audition or a utility like SoX. MakeIR will allow manual creation of multi-channel files, but may well be impractical if you have massive impulse response files.

Avi: If you email me your PCM files (jrp at dial dot pipex dot com), I will have a look at them. (Simple French messages are fine by me.)

Henrik, Thanks. Helpful.

convolver will soon also be available in a DirectShow filter wrapper, which should make it easier to use in DScaler5, ffdshow, etc, and, with suitable impules responses, it should be possible to stick it in a graph that does more interesting things.

Version 1.24 is now available on http://convolver.sourceforge.net.

It includes ConvolverWrapper, which allows convolver to be used as a DirectShow filter. Many thanks to John Adcock.

Thanks. I will try to post instructions for ZoomPlayer once I have worked out the right steps to get ConvolverWrapper to be inserted.

I'd be interested in suggestions for making the plug-in easier to use for multi-channel processing without having to create a single .wav impulse response file.

The main issue is producing a comprehensible interface that covers all the different possible combinations.

I've put up a slightly faster version (1.25).

OK, let's have a go at drawing some of this together.

The current version of convolver takes a single (possible multi-channel) filter file and convolves it with the input stream, channel by channel.

The convolution is probably as fast as it's going to get without moving to another FFT package. On a 2GHz Pentium M or a 3.2GHz Pentium 4 and a long filter, convolver convolves at roughly 20 times real time speed (1.5s to convolve 2 x 65536 x 10 x 44.1KHz samples) -- calculate optimum attenuation convolves 10 filter-lengths' worth of noise, if you would like to report your experience. There are currently two builds up (VC++ 2003 and Intel C++). I'll see if VC++ 2005 or FFTW are faster in due course, but the results on the net are equivocal. Can anyone report how this compares with the likes of BruteFir, or

There is clearly a demand for a more flexible approach.

I suggest the following and would welcome comments:

- allow the user to specify a list of mono impulse files (of equal length, or zero-pad to the maximum length?)

- for each filter allow the user to specify a list of (up to eight?) input channels and scaling factors to which the filter will be applied.

- for each filter, input and scaling factor combination, allow the user to specify an output channel, to which the convolved output will be sent. Where an output file has more than one input filter and scaling factor, sum the scaled outputs

- where an output channel has no convolved input specified, introduce a delay to sync with the convolved outputs

Questions:

- would such a framework be flexible enough to do the job?

- Have I missed any considerations?

- would it be preferable to have the plug-in generate its own filters (eg, high, low mid-band pass, rather than read user-provided filters)?

What would those filters be?

This would obviously not be suitable for room correction, or generate specific cross-overs for their particular setup, but would make it easier for users who don't want to have to generate their own filters. Perhaps the answer is to do both, but in separate plug-ins, to avoid having an overly-complex interface.

- is there a requirement to be able to use stereo impulse files? While theoretically, you should record mono responses, it has been suggested to me that a stereo response file may sound good in some circumstances.

Over to you.

I have now had a report that 1.26 works with ffdshow using the following graph:

Source -> avi splitter -> ffdshow audio dec -> Convolver wrapper -> ffdshow
audio processor -> Direct Sound

cpu usage was 60-80% using 71330 taps @48Khz with 6 channels on a Semprom 3000+

1.26i may perform even better on more recent processors. Processors with larger caches may also perform better.


I am starting work on a version of the plug-in that will allow cross-channel processing (for bass management, etc).

Just looking at how best to design / optimize the cross-channel version:

Looking at the full design sketched above allows the possibility of

- associating a filter with the weighted sum of a specified set of input channels
- summing the weighted outputs of such filters to generate the output for a channel

* Do people need the full generality of such a design?

For example, it would be possible to:

- require that both input and output channels are associated with at most one filter

- require that input channels are associated with at most one filter

- require that output channels are associated with at most one filter; or

- allow only the inputs (or the outputs) to be weighted before they are summed.

While it would be easy to say that "we need the full generality" there will be usabilty / configuration complexity / performance downsides.

What does the panel think?

Well we now have a prototype of a more fully-featured version of the convolution plug-in. It now provides channel-mixing/splitting capabilities:

To summarise:

- Version 1 required an impulse to be presented to it as a multi-channel WAV file, which it convolved, channel by channel, with the input.

- Version 2 requires a set of single-channel impulse response (FIR filter) files. Each impulse response is convolved with the pointwise weighted sum of a specified set of input channels. The resulting outputs are also weighted, summed and directed to specified output channels. So as well as room correction, you should be able to use high and low pass filters, for example, to split an input channel and redirect the output to different output channels. At present, the filter files, weightings and channel redirection are specified through a simple text configuration file. Once the design has been tested further, a user interface will be provided.

The plug-in can also be used as a DirectShow filter using the included ConvolverWrapper, so you should be able to plug it into ffdshow, etc.

Check out the details at convolver.sourceforge.net.

Version 2 is unoptimized -- optimizations will depend in good part on what sort of configurations people use.

Feedback welcome, ideally with the configuration and filters that you are using. Testing of version 2.1 has been pretty limited. jrp at dial dot pipex dot com

Enjoy