Try use copy-vector instead of copy-feats when you are using kaldi.
The error message actually saying Expected token Feature Matrix, got Feature Vector instead
Kaldi defines its own binary format, ark (archive format), this article explains how to read and write or convert this format.
Read/write Kaldi features
We introduce the reading and writing of two type of features here, matrix like MFCC feature, and vector based iVector and VAD vector.
Read Kaldi MFCC features
Raw MFCC Feature location
Most Kaldi features are stored in mfcc folder with an extension of ark or scp.
The ark file
The ark stores raw features, its size is normally in few hundred MBs.
Eg: 20 dimensional MFCC features matrix is stored in the ark file like following:
UtteranceID1 [ d1 d2 d3 d4 d5 .. d20\n d1 d2 d3 d4 d5 .. d20\n d1 d2 d3 d4 d5 .. d20\n ...]\n
UtteranceID2 [ d1 d2 d3 d4 d5 .. d20\n d1 d2 d3 d4 d5 .. d20\n ]\n
Where \n means new line.
To view raw feature, type the following command in the terminal
copy-feats ark:./abc.ark ark,t:
This command means copy the feature from input source (ark:source) to output target (ark,t:target),here we leave “target” empty so the feature will print to the terminal. Following command will dump the features to text file
copy-feats ark:./abc.ark ark,t:a.txt
And dump to binary file:
copy-feats ark:./abc.ark ark:a.bin
The scp file
It is often saw a scp file which describes the content of an ark file.
The scp is only a text file, with following format:
UtteranceID1 arkLocation1:offset1
UtteranceID2 arkLocation2:offset2
Following two commands will give same results
copy-feats scp:./abc.scp ark,t:yourfilename
copy-feats ark:./abc.ark ark,t:yourfilename
Features in the data folder
feats.scp and vad.scp are two feature descriptors in the Kaldi data folder.
Write kaldi MFCC features
One can write kaldi feature to the ark follow the given text format. However most script in Kaldi require its scp file, one way to create scp file is:
copy-feats ark:./abc.ark ark,scp:b.ark,b.scp
Read/write kaldi iVector or VAD vector
Both iVector and VAD vector are in vector form. Read/write kaldi iVector is similar to read/write MFCC feature, only replace the copy-feats with copy-vector command.
Raw iVetor Feature location
Most Kaldi iVector are stored in “exp” folder with an extension of ark or scp.
The VADs are stored in ‘mfcc’ folder.
Read Kaldi iVector
iVector is sorted in the ark in a vector format, please notice that there is a ‘SPACE’ after ‘[’ and no ‘SPACE’ in UtteranceID:
UtteranceID1 [ d1 d2 d3 d4 d5 .. d400 ]\n
UtteranceID2 [ d1 d2 d3 d4 d5 .. d400 ]\n
VAD have same format except the length of VAD is depends on the length of utterance. 1 for voiced frame , 0 for unvoice frame. 1 frame= 10ms
Similar to read MFCC feature,following command will read the iVector:
copy-vector ark:./ivector.1.ark ark,t:YOUR_TEXT_FILE
Write Kaldi iVector
Once you know the iVector format, you can write it back easily. Following command will convert your ark file to Kaldi format, and generate a scp file.
copy-vector ark,t:./your.ark ark,scp:kaldi.ivectors.ark,kaldi.ivectors.scpFAQ:
Where is the copy-feats and copy-vector ? It is in your kaldi src folder. You need include their directory in your path. Please refer to the path.sh in the this SRE folder.
To load the setting in the path.sh, type this in terminal in SRE2010 folder
. path.sh
Or you can put the setting in ~/.profile or ~/.bashrc
Reference:https://wsstriving.github.io/2016/04/17/how-to-check-kaldi-ark/
https://github.com/StevenLOL/Research_speech_speaker_verification_nist_sre2010/blob/master/doc/help_kaldi.md
No comments:
Post a Comment