Hao Tang
@larryniven4.bsky.social
Lecturer at the University of Edinburgh. Member of Centre of Speech Technology Research (CSTR).
Back then Kaldi hadn't existed yet. HTK was still the dominant toolkit. I cannot remember if there was MLP implemented in HTK. I was mostly using a neural network toolkit (I think from ICSI) that just trained 3-layer MLPs. (Do people still have the source code? Somebody should put it on github.)
December 4, 2024 at 11:47 AM
Back then Kaldi hadn't existed yet. HTK was still the dominant toolkit. I cannot remember if there was MLP implemented in HTK. I was mostly using a neural network toolkit (I think from ICSI) that just trained 3-layer MLPs. (Do people still have the source code? Somebody should put it on github.)
Hybrid (Bourlard and Morgan, 1989) and tandem (Hermansky et al., 2000) approaches were well established already. I think the tandem approach had an upper hand, until it was convincingly displaced in 2012. (But hey, we are using HuBERT features nowadays, not very different from a tandem approach.)
December 4, 2024 at 11:47 AM
Hybrid (Bourlard and Morgan, 1989) and tandem (Hermansky et al., 2000) approaches were well established already. I think the tandem approach had an upper hand, until it was convincingly displaced in 2012. (But hey, we are using HuBERT features nowadays, not very different from a tandem approach.)
People were very skeptical, mainly because nobody knew where to even start to reproduce the results. (The results were finally reproduced years later in Kaldi.)
December 4, 2024 at 11:47 AM
People were very skeptical, mainly because nobody knew where to even start to reproduce the results. (The results were finally reproduced years later in Kaldi.)