config.json has no decoder_alias_free — is the shipped decoder plain HiFi-GAN?
Thanks for releasing this. While warm-starting Micro-v2 with the public finetune/ toolkit I ran into something I'd like to confirm before drawing conclusions from it.
The model card and the repo docs describe an "alias-reduced 24 kHz decoder".
But in the released checkpoint:
- config.json has no decoder_alias_free key in its model block, so Generator.init in runtime/models.py falls back to the default False.
- model.pth contains 231 dec.* tensors, and none of them are SnakeBeta alpha/beta or anything from AliasFreeActivation1d / AliasFreeResBlock1 — only conv_pre, ups.* (weight_g/weight_v), resblocks.*, conv_post.
That reads as a standard HiFi-GAN V1 generator (ConvTranspose1d strides 8,8,2,2 with kernels 16,16,4,4, LeakyReLU, ResBlock1), with runtime/inflect_alias_free.py present but inactive for this checkpoint.
Two questions:
- Is that intentional for the v2 release — i.e. is the alias-free path something the runtime supports but the shipped checkpoints do not use, and the "alias-reduced" wording refers to something else?
- If the released decoder was in fact trained with decoder_alias_free=True, could the missing config.json key and the absent Snake parameters be an export issue?
Context for why I'm asking: adapting Micro-v2 to a high-pitched voice (~355 Hz median, vs the release voice around 105 Hz) produces a strong steady comb at exact multiples of 93.75 Hz (= 24000/256, the decoder frame rate) and 750 Hz, present even in silent frames, at 15–20 dB above the local spectrum. Real recordings and the unmodified release rendering English are clean at those frequencies. Anti-imaging in the decoder is one of the things I'm trying to rule in or out, so knowing the shipped configuration is correct matters.