Summary of Kaldi for ivector and xvector

Files List

ivector/

• conf/: configure file for mfcc and vad

• wav/: test audio (you can also use your own wav path, see Step 1)

  • Only supprot flac (install flac), wav and sph (install sph2pipe )

• model_3000h/: pre-trained model

• enroll.sh: main process fille

• data/: save extracted features (It's a generated file)

  • utt2spk, wav.scp generate two files through make_data.py
  • spk2utt: generate from utt2spk
  • log/: save all logs
  • tmp/: save all tmp files

xvector/

• conf/: configure file for mfcc and vad

• wav/: test audio (you can also use your own wav path, see Step 1)

  • Only supprot flac (install flac), wav and sph (install sph2pipe )

• exp/: pre-trained model

• enroll.sh: main process fille

• data/: save extracted features (It's a generated file)

  • utt2spk, wav.scp generate two files through make_data.py
  • spk2utt: generate from utt2spk
  • log/: save all logs
  • tmp/: save all tmp files

format_norm.py: change ark format to npz format

Extract features: ivector and xvector

Step 0: Preparation

• First, install Kaldi.

• Then, step into ivector/ or xvector/ folder

• Change KALDI_ROOT in path.sh to your own kaldi root

• Add link:

ln -s $KALDI_ROOT/egs/sre16/v2/steps ./
ln -s $KALDI_ROOT/egs/sre16/v2/sid ./
ln -s $KALDI_ROOT/egs/sre16/v2/utils ./

Step 1: Extract ivector and xvector

Refers to pre-trained xvector model in kaldi and kaidi-sre-code

• Extract ivector: cd ivector and run enroll.sh to extract ivector

bash enroll.sh wav_path
# for example: bash enroll.sh ./wav

• Extract xvector: cd xvector and run enroll.sh to extract ivector

## Case 1: extract xvector without speaker infos
#          for example: bash enroll.sh ./wav 1
bash enroll.sh wav_path 1

## Case 2: extract xvector with speaker infos
####Step 1: Generate speaker.txt files
####Only suitable for files like 'wav_root/speaker_id/wav_name'
####Other format, you should write your own generate_speaker.py
python generate_speaker.py wav_dir speaker.txt
####Step 2: extract xvector with speaker infos
####For example, bash enroll.sh ./speaker.txt 2
bash enroll.sh ./speaker.txt 2

Step 2: Read generate ivector and xvector

In this section, we convert ivector and xvector from ark type to array type

i-vector in data/feat/ivectors_enroll_mfcc

• spk_ivector.ark i-vector for each speaker
• ivector.1.ark: i-vector for each utturance (400-d i-vector)

x-vector in data/feat/xvectors_enroll_mfcc

• spk_xvector.ark x-vector for each speaker
• xvector.1.ark: x-vector for each utturance (512-d x-vector)

## print name and feats from ark to txt
$KALDI_ROOT/src/bin/copy-vector ark:ivector/data/feat/ivectors_enroll_mfcc/ivector.1.ark ark,t:- >ivector.txt

$KALDI_ROOT/src/bin/copy-vector ark:xvector/data/feat/xvectors_enroll_mfcc/xvector.1.ark ark,t:- >xvector.txt

## Or you can change ark format to np.array format, which has (data_path ['pic_path'], ivector or xvector)
python format_norm.py --vector_path='xvector.txt' --save_path='x_vector.npz'
python format_norm.py --vector_path='ivector.txt' --save_path='i_vector.npz'

Other summary

## combine different files
utils/combine_data.sh

## make xxx fits to the kaldi format
utils/fix_data_dir.sh xxx

## gain subset of data
utils/subset_data_dir.sh

## file exists and dir exists
if [ -d "./data" ];then # dictionary exists
if [ -f "./data/1.txt" ];then # file exists

## xvector/run.sh
Has four folder: 
	sre_combined (source domain, argument data, for training)
	sre16_major (unlabeded target domain for model adaption)
	sre16_eval_enroll(labeded target domain for train)
	sre16_eval_test(unlabeded target domain for test)
Main stream: xvector->mean->transform(LDA)->len normalize->classifier(PLDA/adapt-PLDA)

## ark: split by space and print the third one
echo '1 2 3' |awk '{print $3}'  # print 3