Skip to content

Latest commit

 

History

History
34 lines (23 loc) · 1.07 KB

readme.md

File metadata and controls

34 lines (23 loc) · 1.07 KB
Raw

Extract Speech Features Using Kaldi

After updating the paths in these files, run:

source cmd.sh
source path.sh

Prepare format for Kaldi feature extraction:

./flickr8k_data_prep_vad.py

Extract filterbank features:

./steps/make_fbank.sh --compress false --nj 30 --cmd "$train_cmd" \
    data/full_vad exp/make_fbank/full_vad fbank
cat fbank/raw_fbank_full_vad.*.scp > fbank/fbank_full_vad.scp

The --nj argument specifies the number of CPU cores to use.

Extract MFCC features:

./steps/make_mfcc.sh --nj 30 --cmd "$train_cmd" data/full_vad \
    exp/make_mfcc/full_vad mfcc
./steps/compute_cmvn_stats.sh data/full_vad exp/make_mfcc/full_vad mfcc
./local/cmvn_dd.sh --nj 30 --cmd "$train_cmd" data/full_vad \
    exp/make_cmvn_dd/full_vad mfcc
cat mfcc/mfcc_cmnv_dd_full_vad.*.scp > mfcc/mfcc_cmnv_dd_full_vad.scp

Get duration of complete corpus:

cat data/full_vad/segments | \
    awk 'BEGIN {sum = 0.0} {sum += $4 - $3} \
    END {printf "Total duration: %.6f hours\n", sum/60.0/60.0}'