Macbook M1 PRO (extremely slow!)

[dxcore35]

As there is only CPU support for non-nvidia folks. I tried to create small audio from textfile

Category	Details
HARDWARE	MacBook M1 Pro Max
SOFTWARE	Docker
INPUT FILE	TXT file - 11K words
CPU Usage	>10%
TIME	❗️ Running for 10 minutes - Only 12% done

I don't know but this is is completely useless to use it on Mac OS.

[ROBERT-MCDOWELL]

So sell your Mac and buy a PC with nvidia card... frankly this is not an ebook2audiobook issue but all A.I libraries we are using for the speech, which is tensorflows, torch, coqui-tts etc... and M1 is well known to not supported by any of them....

[dxcore35]

I checked and there is support for Mac:

Native Apple Silicon (M1/M2) Support for TensorFlow, PyTorch, and Coqui-TTS

Library	Native Support	Installation Commands	GPU Support	Metal API Backend Details	GPU Verification
TensorFlow	✅ Yes (2.5.0+)	pip install tensorflow-macos tensorflow-metal	Enabled via tensorflow-metal	Optimized with Metal Performance Shaders (MPS)	import tensorflow as tf print(tf.config.list_physical_devices('GPU'))
PyTorch	✅ Yes (1.12.0+)	pip install torch torchvision torchaudio	Enabled via MPS backend	Supports Metal via Metal Performance Shaders (MPS)	import torch print(torch.backends.mps.is_available())
Coqui-TTS	✅ Yes	pip install TTS (Requires TensorFlow or PyTorch installation)	Depends on TensorFlow/PyTorch GPU	Relies on TensorFlow/PyTorch for Metal integration	Same as TensorFlow/PyTorch verification based on the underlying library used.

---

Notes

• Metal Support: Apple’s Metal API is used to enable GPU acceleration for deep learning libraries, offering improved performance on Apple Silicon.
• Requirements:
  • macOS 11.0+ (Big Sur or later)
  • Xcode Command Line Tools
• Performance:
  • TensorFlow offers the most mature Metal support.
  • PyTorch’s MPS backend is functional but still evolving.
  • Coqui-TTS performance depends on the framework it's paired with (TensorFlow or PyTorch).

[ROBERT-MCDOWELL]

yes, you can use these libraries on M1, but with your CPU, not GPU unless you have installed an nvidia.
or, if I'm wrong, so modify our code and send a PR.

[ROBERT-MCDOWELL]

for now what we can do is to check if MPS is available, it will be more optimized but will still work on CPU
While MPS can speed up inference, training performance might still lag compared to CUDA.

[ROBERT-MCDOWELL]

Added "mps" device for the next git update/release, thanks to test if it will work for you.

[dxcore35]

For tensorflow-macos

TensorFlow on Apple Silicon utilizes the CPU’s multiple cores (using the tensorflow-macos version) without needing NUMA-like handling, as the system’s unified memory allows for more efficient data sharing between cores.

[dxcore35]

For torch

I tried:

import torch

# Check MPS availability
if torch.backends.mps.is_available():
    device = torch.device("mps")  # Use GPU
    print("Using MPS (GPU):", device)

    # Example tensor on GPU
    x = torch.randn(3, 3).to(device)
    print("Tensor on GPU:", x)
else:
    print("MPS backend is not available. Running on CPU.")

I got:

Using MPS (GPU): mps
Tensor on GPU: tensor([[ 0.4005, -0.1923, -0.5197],
        [ 1.2033,  0.5933,  1.6042],

[ROBERT-MCDOWELL]

torch is one thing, coqui-tts is another thing using only a a part of torch, depending its version etc...
nothing is easy in A.I. development and it's not a simple test with one library that will solve everything.
as I said (and you must read my comment above), a patch will be done for the next update.
so if you are ok to test once updated I live this issue open, if not, I close it.

[dxcore35]

Yes I read it. I know the Apple silicon is for some reason not compactible with multiple libraries for AI. I struggle to find some TTS that will run on this hardware...

Yes can test it on same file, just please let me know when new version will be ready. I will test it and post the difference in speed.

[ROBERT-MCDOWELL]

I don't know when I will update as I'm stuck with other issues for now... follow the repo and you'll receive an email when released.

[landerhe]

don't think about adding MPS support just yet, it will face a issue with the tts model, MPS does not support convolution operations with output channels greater than 65536. It's definitely not a priority to-do but one that will require loads of look into.

[DrewThomasson]

yup, as seem here xtts def has issues trying to run on mps

idiap/coqui-ai-TTS#65

[ROBERT-MCDOWELL]

nevermind, it's added already so maybe coqui-tts fork guys will fix it in the near future.