iamyuanchung / autoregressive-predictive-coding Goto Github PK

I trained APC by myself data process , but it seem don't convergence in train. and your dataset is too slow to download . I want to know what's the value of loss when you stop your training ?thanks

Hi,
Thank you for sharing your great work!
It seems that the Dropbox links for the dataset and pre-trained weights are no longer valid (they give "Something went wrong" error).
Could you please update these links?
Thanks!

Hi, Yu-An Chung,
Thank you for releasing your work on APC. I am doing reseach on topic music information retrieval and think your work may helpful for us.
I found that Transformer-APC is not included in current code version. So as to the test part for the ASR test experiment in paper GENERATIVE PRE-TRAINING FOR SPEECH WITH AUTOREGRESSIVE PREDICTIVE CODING.
I wonder whether you will have plan to release them out in the future? Hope to get your reply.

Hi,
There's a problem with the feature files.

After downloading the train-clean-360.xz and dev-clean.xzvia dropbox link on the README.md,
I've tried to unzip them

But xz -d just produced train-clean-360 (no file extensions), which is not readable with numpy or pickle.

Can I get the feature files?

Thank you :)

Hi, do you have validation loss scores for all models (n=1,...n=20) on libri-valid-clean?
I want to verify whether my results are correct. I created a test script and loaded your pre-trained models.

I got these validation losses:
n=1, 0.30188
n=3, 0.50266

Thank you in advance.

Hi,
Thank you for sharing your code. Great work !
Since I think preprocessing parameters like hop length and window length are important for APC. Also I want to do experiments on other dataset.
So could you please share them, or share a preprocessing script ?

Hi, can you also share your preprocessing script to generate 80-dimensional log Mel spectrograms? Did you apply the same script to both Librispeech and WSJ? Thank you

I find that the T-APC (transformer-based version APC in GENERATIVE PRE-TRAINING FOR SPEECH WITH AUTOREGRESSIVE PREDICTIVE CODING) is not included in this repository currently.
Any plans to release the T-APC code?
THX!

Hi!
Thanks for the great lib and interesting paper.

What sample rate did you use for loading audio and is this important when computing log mel specs?
Also what is model license?

Thank you kindly :)

Hello, I'm currently trying APC, but I'm not familiar with kaldi, and I found that torchaudio provide fbank function that match Kaldi’s compute-fbank-feat. I'm wondering is it possible to use fbank for creating the same feature?If so, how can I set the parameter?