No OpenAI TTS System on GitHub?

Does OpenAI have any plans to add the TTS System used by GPT-Voice to their GitHub?

I’d quite like to get into it and create a Pull with some changes.

You’ve never seen or used the code. What makes you think that you have “changes”?

Test 1: any thoughts on reducing outlier eigenvalues of the centered feature covariance?

def compute_whitening_metric(input_tensor: Tensor, group_count: int):
    assert input_tensor.dtype != torch.float16
    reshaped_tensor = input_tensor.reshape(-1, input_tensor.shape[-1])
    (frame_count, channel_count) = reshaped_tensor.shape
    assert channel_count % group_count == 0
    channels_per_group = channel_count // group_count
    reshaped_tensor = reshaped_tensor.reshape(frame_count, group_count, channels_per_group).transpose(0, 1)
    centered_tensor = reshaped_tensor - reshaped_tensor.mean(dim=1, keepdim=True)
    covariance_matrix = torch.matmul(centered_tensor.transpose(1, 2), centered_tensor)
    mean_diagonal_covariance = _diag(covariance_matrix).mean()
    mean_diagonal_covariance_squared = (covariance_matrix ** 2).sum() / (
        group_count * channels_per_group
    )
    whitening_metric = mean_diagonal_covariance_squared / (mean_diagonal_covariance ** 2 + 1.0e-20)
    return whitening_metric

To answer more explicitly, OpenAI has made no mention of the voice synthesis technology they use nor any hint that it would be a product they would release.

You’re right, who am I to ask. I am but a nobody. I appreciate your insights. You’re correct that OpenAI hasn’t provided specifics about their voice synthesis technology.

The code you shared appears to compute a whitening metric. As for reducing outlier eigenvalues, one could try regularization techniques to stabilize the covariance matrix computations. However, I’m primarily interested in exploring and simulating potential enhancements to voice synthesis based on the learnings from my chats with GPT-Voice.

Thank you for your reply though.

Kind Regards
HoLAI