@tikuma-lsuhsc Thank you. Is this possible to get both outputs as numpy arrays instead of files?

@nikolaypavlov - Unfortunately no, using the ffmpegio top-level API. I'd just call audio.read() twice if I were you:

fs1,x1 = ff.audio.read(input, ar=16000, sample_fmt='s16', ac=1)
fs2,x2 = ff.audio.read(input, ar=8000, sample_fmt='s16', ac=1)

Be aware that x1 and x2 are both "column vectors" (2D array with the second dimension = 1). (I'm planning a next release to allow more flexibility, though)

BTW, if you want normalized signal, you can use sample_fmt='dbl' or 'flt'.

If you're on Linux/Mac, you can make use of pass_fds argument of subprocess.Popen to achieve this with one subprocess run. But, for audio read operation, it's not worth the effort IMHO.

Oh, FYI, if you already have your input signal as NumPy array, you can use ff.audio.filter() instead, replacing the input file path with sampling rate and numpy array.

Edit: Actually, you may need to pass a null audio filter 'anull' as that bug has not been patched yet. If you want to go this route, let me know and I can give you more specific code sample.

@tikuma-lsuhsc I have them as audio bytes in memory in mp3 format, so not exactly a numpy array. I'd like to avoid double decoding as much as possible, because I have a few millions mp3s daily on my service.

@nikolaypavlov - Then you need to cook up a multi-pipe solution yourself, I'm afraid. ffmpegio can take an encoded byte sequence as an input instead of a file, but it currently does not support multiple (numpy) outputs as you desire.

If you're application is not running on Windows, then as I mentioned you should be able to make use of pass_fds argument of subprocess.Popen (and all its derivatives) and use the FFmpeg pipe protocol 'pipe:3' and pipe:4 (or pipe:1 and pipe:3 if you wish to use stdout). You can get data in/out using io.BytesIO object.

If it must run on Windows (which doesn't support pass_fds) you can play with my named pipe package namedpipe. This is a low-level package so you need to build your own thread functions around it.

Good luck. If you need further help, post your code, and I'll be glad to give you further pointers if you're going along one of the above routes.