make ppgs kaldi implentation
- inputs: 13-dim MFCC + deltas + delta-deltas (39-dim total)
- outputs: 128-dim PPGs
- cd ppgs_maker/timit/s5;
- bash run.sh;
- bash ./local/nnet/run_dnn_deltas.sh;
- cd make_ppgs
- prepare data (wav.scp, spk2utt, utt2spk,...) as in typical kaldi egs;
- get mfcc features: bash ./local/make_mfcc_feats.sh;
- get deltas, delta-deltas features: bash ./local/make_delta_feats.sh;
- get PPGs: bash ./local/run_dnn_forward.sh;
- save PPGs to python dict: bash ./local/make_all_ppgs.sh.
- prepare data;
- modify PPGs dimension by modifing the numLeavesTri1 and numGaussTri1 in ppgs_maker/timit/s5/run.sh, use gmm-info or tree-info tools to check the model's (*.mdl) number of tree leaves.
- modify variables relates to path;
- modify the network structure by:
(1) modifing --nn-depth and --hid-dim in ppgs_maker/timit/s5/local/nnet/run_dnn_deltas.sh, line 58;
(2) modifing the number before .dbn to the number of layers of the DNN;
- run inference using nnet-forward, parameters needed: feature transform, trained nnet model, gpu option.