xminte / yesno_ice_phones

A modified version of the Kaldi-ASR yesno example project adjusted for Icelandic, for multiple speakers, and it uses a phone-based model instead of word based-model.

Kaldi-ASR Advanced Project
T-718-ATSR - Automatic Speech Recognition, TD-MSc, 2019-1
Reykjavik University - School of Computer Science, Menntavegur 1, IS-101 Reykjavik, Iceland

Table of Contents

Click to expand

1. Introduction
2. The Dataset
3. The Adjustments
   • run.sh
   • create_yesno_txt.pl
     • create_yesno_waves_test_train.pl
     • prepare_data.sh
     • mfcc.conf
4. Authors
5. License
6. References

1 Introduction

A modified version of the Kaldi-ASR yesno example project adjusted for Icelandic, for multiple speakers, and finally, it it uses a phone-based model instead of word based-model.

2 The Dataset

Custom yes/no dataset for Icelandic, made from recordings from students and the staff members from the course.

Summary:

• 10 different speakers
• 120 sentences/utterances
• Speaker 0: 10 utterances
• Speaker 1: 7 utterances
• Speaker 2: 8 utterances
• Speaker 3: 10 utterances
• Speaker 4: 10 utterances
• Speaker 5: 42 utterances
• Speaker 6: 0 utterances
• Speaker 7: 10 utterances
• Speaker 8: 8 utterances
• Speaker 9: 10 utterances
• Speaker 10: 5 utterances
• Each utterance consist of 8 words.
• The vocabulary for the words:
• Já (yes)
• Nei (no

3 The Adjustments

3.1 run.sh

Edit following part in the file:

if [ ! -d waves_yesno ]; then
  wget waves_yesno.tar.gz || exit 1;
  tar -xvzf waves_yesno.tar.gz || exit 1;
fi

We have our dataset locally, so we dont have to fetch it from a remote file host.

Remember to re-compile the script.

$ chmod 777 s5/run.sh

3.2 local/create_yesno_txt.pl

Add following code to the while loop:

    $trans =~ s/SA\d\d-//;

This will remove the speaker Id from our filename when creating our data/{X}/text files.

3.3 local/create_yesno_waves_test_train.pl

Specify which speakers you want to have in the test set.

while ($l = <FL>)
{
	chomp($l);
	if (index($l, "SA01") == -1  && index($l, "SA10") == -1)
	{
		print TRAINLIST "$l\n";
	}
	else
	{
		print TESTLIST "$l\n";
	}
}

3.4 local/prepare_data.sh

Edit the code so each utterece has it corresponding speaker id istead of global.

  cat data/$x/text | awk '{printf("%s %s\n", $1, substr($1, 0, 5));}' > data/$x/utt2spk

Remember to re-compile the script.

$ chmod 777 s5/local/prepare_data.sh 

3.5 conf/mfcc.conf

Set the correct sample-frequency configuration based on the waves_yesno .wav file format.

--sample-frequency=16000 #  waves_yesno is sampled at 16kHz

3.6 input/lexicon.txt

For changing the model from a word-based model to a phone-based one, we will need to change the lexicon for our system.

input/lexicon.txt

<SIL> SIL
YES J A U
NO N E I

For simplicity we cept the same labels, but changed the corresponding phonems.

lexicon_nosil.txt

YES J A U
NO N E I

phones.txt

SIL
A
E
I
J
N
U

Done

Now we should be able to run the run.sh file successfully.

$ cd s5/
$ bash run.sh

Expected output result:

%WER 11.46 [ 11 / 96, 7 ins, 1 del, 3 sub ] exp/mono0a/decode_test_yesno/wer_14_1.0

4 Authors

• Egill Anton Hlöðversson - MSc. Language Technology Student

5 License

This project is licensed under the MIT License - see the LICENSE file for details.

6 References

• Kaldi-ASR
• Kaldi-ASR yesno

🌟 PLEASE STAR THIS REPO IF YOU FOUND SOMETHING INTERESTING 🌟