Firstly, I want to map acoustic features to Bottle-neck Features. And then use Bottle-neck Feature to train another DNN. Could you tell me which current script support this scheme?
Thanks in advance!
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
In egs/wsj/s5/run.sh, search for 'bottleneck', there is an example.
It is not as fast as it could be (if using GPUs), it would be better
to update that script to use the online preconditioning, changing
train_tanh_bottleneck.sh to train_tanh_bottleneck_fast.sh as in
train_tanh.sh -> train_tanh_fast.sh.
Dan
Firstly, I want to map acoustic features to Bottle-neck Features. And then
use Bottle-neck Feature to train another DNN. Could you tell me which
current script support this scheme?
I'm trying to figure out how to use the Bottleneck features...
Why do we need the align_dir for the bottleneck training?
what is the input of the trained ANN? It looks like the input is the acoustic features, am i right? is there a way to use the raw data + acoustic features as input?
Is there an example script to how to use BN for GMM-HMM training and decoding?
Thank you very much
Tamir
Last edit: Tamir 2015-07-01
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I'm trying to figure out how to use the Bottleneck features...
Why do we need the align_dir for the bottleneck training?
Any supervised method needs the supervision labels, and align_dir provides that.
and what is the input of the trained ANN? I guess it should be the raw data,
or raw data + PLP (according to literature), but It looks like the train is
with the plp features
Likely the input is the PLP or MFCC features or the log mel filterbank
energies- these are the standard approaches. A few people have
published experiments with using the raw signal but no-one has shown
that this is better than processed features such as PLP.
Dear All,
Firstly, I want to map acoustic features to Bottle-neck Features. And then use Bottle-neck Feature to train another DNN. Could you tell me which current script support this scheme?
Thanks in advance!
In egs/wsj/s5/run.sh, search for 'bottleneck', there is an example.
It is not as fast as it could be (if using GPUs), it would be better
to update that script to use the online preconditioning, changing
train_tanh_bottleneck.sh to train_tanh_bottleneck_fast.sh as in
train_tanh.sh -> train_tanh_fast.sh.
Dan
On Wed, Dec 3, 2014 at 5:44 PM, Lee speechspeech@users.sf.net wrote:
HI, Povey
Thanks for your kindly reply. I will try it.
Hello,
I'm trying to figure out how to use the Bottleneck features...
Why do we need the align_dir for the bottleneck training?
what is the input of the trained ANN? It looks like the input is the acoustic features, am i right? is there a way to use the raw data + acoustic features as input?
Is there an example script to how to use BN for GMM-HMM training and decoding?
Thank you very much
Tamir
Last edit: Tamir 2015-07-01
Any supervised method needs the supervision labels, and align_dir provides that.
Likely the input is the PLP or MFCC features or the log mel filterbank
energies- these are the standard approaches. A few people have
published experiments with using the raw signal but no-one has shown
that this is better than processed features such as PLP.
Dan