search

mct10 / RepCodec

Models and code for RepCodec: A Speech Representation Codec for Speech Tokenization
Other
161 stars 10 forks source link

About the settings in training Decoder-only ASR model. #10

Open cwzhang98 opened 6 days ago

cwzhang98 commented 6 days ago

Hi Chutong: Do you initialize the embedding layer of the ASR model with VQ codebook parameters? Or just use the semantic tokens and then randomly initialize the embedding layer.

HuangZhiChao95 commented 6 days ago

We randomly initialize the embedding layer with the fairseq library.