support loading audio as float16

[sammlapp]

Hi, it would be great to support loading in float16 dtype. Many autonomous audio recorders used for wildlife monitoring record very long, float16 files, and loading them as float32 doubles the memory usage unnecessarily. Currently, only ['float32', 'float64', 'int16', 'int32'] are supported by sf.load()

It seems like the crux is that ctypes does not have float16

python-soundfile/soundfile.py

Line 149 in 3503941

_ffi_types: Final[dict[str, str]] = {

the audio frame loading happens in

python-soundfile/soundfile.py

Line 1426 in 3503941

def _cdata_io(self, action, data, ctype, frames):

which is called by SoundFile.read() via :

python-soundfile/soundfile.py

Line 1412 in 3503941

def _array_io(self, action, array, frames):

Perhaps it would be possible to cast frames to float16 immediately when loading into python, to avoid the memory overhead of creating the entire array as float32?

[sammlapp]

Update: I was wrong, the common format for ARU is PCM_16. So I can simply load it as int16 then divide by 32768 to get the float value, assigning it to a float16 array. This results in needing to make an in-memory copy though, so peak memory usage is still 4 bytes per sample (same as loading float32).

[bastibe]

This would require float16 support in libsndfile, which did not last time I checked.