Aaron Keesing

@pourfard You should change the title of this issue and preface with 'Request:' so that people will not be confused and think you're providing a build.

For Linux Python 3.8: [tensorflow_cpu-2.4.0-cp38-cp38-linux_x86_64.whl](https://github.com/agkphysics/tensorflow-wheels/releases/download/tf_2.4.0_nogpu_noavx_nomkl/tensorflow_cpu-2.4.0-cp38-cp38-linux_x86_64.whl)

It was built with `-march=x86_64`, so no AVX, AVX2, or SSE4.

> @agkphysics Can you please help me? Unfortunately I have never built TF on Windows and I suspect it will be more difficult to set up the build process than...

Now that https://github.com/audeering/opensmile/issues/20 is fixed, can this update be propagated to the python module?

I can confirm that loading is significantly slower in 0.9.2, even with resampy 0.4.0. I have an audio file which is ~2 sec and 60 KB. The below code reports...

I believe the issue may be with calling `audioread.available_backends()`: https://github.com/librosa/librosa/blob/b9830b4a335b1e7e4be66ee81870573bf3d516d0/librosa/core/audio.py#L158 A quick test in IPython shows this call is adding ~50 ms to each call of `load()`: ```ipython In [1]:...

Yes I can confirm that with `audioread==3.0.0` the audio loading times are the fast for both `librosa==0.9.2` and `librosa==0.9.1`.

I've also encountered this issue with unlimited dimensions but I solved it similar to https://github.com/Unidata/netcdf4-python/issues/859, by increasing the chunksize of the unlimited dimension.

@fehrin I will try and build with py39 and py310 when I get time. @dathudeptrai This wheel is built with TensorRT 8.4.3 but I'm not sure how close it is...