Request: trainable checkpoints (Lightning .ckpt / state_dicts) for finetuning
Hello, and thank you for releasing Sidon — the restoration quality is excellent and the 500x RTF claim holds up in our testing.
Use case: We are building a large open Quranic recitation dataset (thousands of hours, in-the-wild mosque/radio recordings) for TTS and tajweed-analysis research. Sidon v0.1 already performs impressively on this material, but Quranic recitation is heavily out-of-domain relative to typical multilingual ASR corpora: very long sustained vowels (madd, up to several seconds), pharyngeal/emphatic consonants, and characteristic large-room reverb. We would like to domain-finetune Sidon on a curated ~100h clean recitation pool using your released training code and augmentation pipeline.
The blocker: the v0.1 release ships TorchScript traced graphs (feature_extractor_cuda.pt, decoder_cuda.pt), which are inference-only — the traced graphs cannot be loaded back into the src/sidon/model Lightning modules for continued training.
Request: would you be willing to release (or share) either:
- the Lightning checkpoints (
last.ckptor equivalent) for the w2v-BERT feature predictor and the vocoder, or - plain
state_dicts compatible with the model classes insrc/sidon/model?
Failing that, any guidance on the exact w2v-BERT 2.0 finetuning configuration you used (so we can reproduce the predictor from the public base) would be greatly appreciated.
We would of course share the resulting recitation-domain checkpoint back with the community under the same spirit as your release. Thank you for the great work!
Update: we found sarulab-speech/sidon_raw_weight with the LoRA adapter and decoder state dict - exactly what we needed for finetuning, thank you for publishing those. One remaining ask: would you consider also releasing the GAN discriminator weights from the vocoder finetune stage? For continued finetuning (in our case, adapting Sidon to Quranic recitation), restarting the discriminator from scratch risks degrading the pretrained generator early in training. Even an un-cleaned Lightning checkpoint would be very useful.
(The comment above was posted from a colleague's account by mistake - this is the right one.)