Creating the acoustic model yourself

From RobotinoWiki

Setting up Speech Recognition

Setting up speech recognition is done in two parts. The first part involves training of the speech data and creation of the acoustic model and the second involves the actual execution of the speech recognition engine with the created acoustic model.

Creating an Acoustic Model

The HTK toolkit (version 3.4) shall be used to create an acoustic model. Please follow the following steps.

Step 1: Register with HTK

You will need to register with HTK before you can download it. Please do so.

Step 2: Download HTK sources

Download the sources for HTK toolkit 3.4 from here. Also download the HTK samples from here.

Step 3

  • Move to your home directory
 cd ~ 
  • Create a directory called 'bin'
 mkdir bin 
  • Unpack the downloaded HTK toolkit sources and HTK samples in a folder called 'htk-3.4' in the 'bin' directory. The 'bin' directory should contain the following
  
htk-3.4 samples
  • Move the 'samples' folder to the 'htk-3.4' folder as follows
cd bin
mv samples htk-3.4
  • If you have a newer version of the gcc compiler (version 4 or above), you will need to install gcc version 3.4 so that HTK will compile properly. Use the following gcc's version command to see which version is installed on your system
gcc -v
  • If your gcc version is 4 and above, follow the listed commands to install gcc 3.4
 
sudo apt-get install gcc-3.4 
sudo rm /usr/bin/gcc
sudo ln -s /usr/bin/gcc-3.4 /usr/bin/gcc 

NOTE - if the above doesn't work for you then maybe the hardy ubuntu package repository is not in your sources.list file. In that case, do the following. If it does work, then skip to the next bullet point.

sudo gedit /etc/apt/sources.list
add the following line to the end of the sources.list file
deb http://de.archive.ubuntu.com/ubuntu/ hardy main universe
now run the following command
sudo apt-get update
This should pull the hardy ubuntu packages from the repository. You can now run the following commands.
 
sudo apt-get install gcc-3.4 
sudo rm /usr/bin/gcc
sudo ln -s /usr/bin/gcc-3.4 /usr/bin/gcc 
  • Install the external dependencies as follows
sudo apt-get install libx11-dev libesd0-dev libasound2-dev libzip1 flex libncurses-dev
  • Now move to the 'htk-3.4' dir and configure htk as follows. Note change %yourusername% from the command to your user name.
cd htk-3.4
./configure --prefix=/home/%yourusername%/bin/htk-3.4
  • Now run make all and make install. This should install the created binaries to the folder '/home/yourusername/bin/htk-3.4/bin' .
make all
make install
  • Change directory back to home and create a folder called 'voxforge' and then a folder called 'HTK_scripts' in the voxforge folder.
cd ~
mkdir voxforge
cd voxforge
mkdir HTK_scripts
cd HTK_scripts
  • Now copy some scripts from the 'htk-3.4/samples' folder to the 'HTK_scripts' folder as follows
cp ../../bin/htk-3.4/samples/RMHTK/perl_scripts/mkclscript.prl .
cp ../../bin/htk-3.4/samples/HTKTutorial/maketrihed .
cp ../../bin/htk-3.4/samples/HTKTutorial/prompts2mlf .
cp ../../bin/htk-3.4/samples/HTKTutorial/prompts2wlist .
  • Your 'HTK_scripts' folder should contain the following
maketrihed  mkclscript.prl  prompts2mlf  prompts2wlist

Step 4

  • Now we will download Julius (version 4.5.1). We shall be using pre-compiled binaries which can downloaded from here
  • Once downloaded extract them to your '/home/%yourusername%/bin' folder. After that is done your 'bin' folder should contain the following
htk-3.4  julius-4.1.5-linuxbin

Step 5

  • Now you will need to update your user path which can be done as follows. First change to your home directory and edit the .bashrc file.
cd ~
gedit .bashrc
  • Add the following to the end of the .bashrc file. Note change %yourusername% from the command to your username.
# HTK and JULIUS scripts and executables
PATH=$PATH:$HOME/bin:/home/%yourusename%/bin/htk-3.4/bin:/home/%yourusename%/bin/julius-4.1.5-linuxbin/bin
  • Source your .bashrc file to reflect the changes
source ~/.bashrc
  • Test if your HTK toolkit has been installed correctly by running the following command.
HVite -V
You should see an output similar to the following.

HTK Version Information
Module     Version    Who    Date      : CVS Info
HVite      3.4        CUED   25/04/06  : $Id: HVite.c,v 1.1.1.1 2006/10/11 09:55:02 jal58 Exp $
HShell     3.4        CUED   25/04/06  : $Id: HShell.c,v 1.1.1.1 2006/10/11 09:54:58 jal58 Exp $
HMem       3.4        CUED   25/04/06  : $Id: HMem.c,v 1.1.1.1 2006/10/11 09:54:58 jal58 Exp $
HLabel     3.4        CUED   25/04/06  : $Id: HLabel.c,v 1.1.1.1 2006/10/11 09:54:57 jal58 Exp $
HMath      3.4        CUED   25/04/06  : $Id: HMath.c,v 1.1.1.1 2006/10/11 09:54:58 jal58 Exp $
HSigP      3.4        CUED   25/04/06  : $Id: HSigP.c,v 1.1.1.1 2006/10/11 09:54:58 jal58 Exp $
HWave      3.4        CUED   25/04/06  : $Id: HWave.c,v 1.1.1.1 2006/10/11 09:54:59 jal58 Exp $
HAudio     3.4        CUED   25/04/06  : $Id: HAudio.c,v 1.1.1.1 2006/10/11 09:54:57 jal58 Exp $
HVQ        3.4        CUED   25/04/06  : $Id: HVQ.c,v 1.1.1.1 2006/10/11 09:54:59 jal58 Exp $
HModel     3.4        CUED   25/04/06  : $Id: HModel.c,v 1.2 2006/12/07 11:09:08 mjfg Exp $
HParm      3.4        CUED   25/04/06  : $Id: HParm.c,v 1.1.1.1 2006/10/11 09:54:58 jal58 Exp $
HDict      3.4        CUED   25/04/06  : $Id: HDict.c,v 1.1.1.1 2006/10/11 09:54:57 jal58 Exp $
HNet       3.4        CUED   25/04/06  : $Id: HNet.c,v 1.1.1.1 2006/10/11 09:54:58 jal58 Exp $
HRec       3.4        CUED   25/04/06  : $Id: HRec.c,v 1.1.1.1 2006/10/11 09:54:58 jal58 Exp $
HUtil      3.4        CUED   25/04/06  : $Id: HUtil.c,v 1.1.1.1 2006/10/11 09:54:59 jal58 Exp $
HAdapt     3.4        CUED   25/04/06  : $Id: HAdapt.c,v 1.2 2006/12/07 11:09:07 mjfg Exp $
HMap       3.4        CUED   25/04/06  : $Id: HMap.c,v 1.1.1.1 2006/10/11 09:54:57 jal58 Exp $

  • Test if Julius has been installed correctly by entering the following command in the terminal
julius-4.1.5
You should see an output similar to the following
Julius rev.4.1.5 - based on 
JuliusLib rev.4.1.5 (fast)  built for i686-pc-linux

Copyright (c) 1991-2009 Kawahara Lab., Kyoto University
Copyright (c) 1997-2000 Information-technology Promotion Agency, Japan
Copyright (c) 2000-2005 Shikano Lab., Nara Institute of Science and Technology
Copyright (c) 2005-2009 Julius project team, Nagoya Institute of Technology

Try '-setting' for built-in engine configuration.
Try '-help' for run time options.
  • Now to switch back to your original gcc version, do the following (The original version in my case was 4.3, yours may differ)
 
sudo rm /usr/bin/gcc
sudo ln -s /usr/bin/gcc-4.3 /usr/bin/gcc 

Step 6

Install Audacity as follows

sudo apt-get install audacity

Step 7

We will now compile the grammar and voca files.

  • Create a folder called 'auto' in your '/home/%yourusename%/voxforge' directory
cd ~
cd voxforge
mkdir auto
cd auto
  • Download the robotino.grammar and robotino.voca files and extract them in your 'auto' folder you just created. After extraction your 'auto' folder should contain the following
robotino.grammar  robotino.voca
  • Now compile the grammar and voca files to Julius files. Make sure you are in the 'auto' folder. Run the following command
mkdfa.pl robotino

KNOWN ERROR - in case you get an error as follows while running the command above

/usr/X11R6/bin/perl: bad interpreter: No such file or directory
Then open the mkdfa.pl file
gedit ~/bin/julius-4.1.5-linuxbin/bin/mkdfa.pl
And change the first line from
#!/usr/X11R6/bin/perl
To
#!/usr/bin/perl
And run the command again
mkdfa.pl robotino
You should see an output as follows
robotino.grammar has 12 rules
robotino.voca    has 12 categories and 29 words
---
Now parsing grammar file
Now modifying grammar to minimize states[6]
Now parsing vocabulary file
Now making nondeterministic finite automaton[34/34]
Now making deterministic finite automaton[29/29] 
Now making triplet list[29/29]
12 categories, 29 nodes, 38 arcs
-> minimized: 15 nodes, 24 arcs
---
generated: robotino.dfa robotino.term robotino.dict

Step 8

Now we shall proceed to the training and creation of the acoustic model.

  • Download the prompts file and the codetrain.scp and save it in your '/home/%yourusername%/voxforge/auto' folder. Your 'voxforge/auto' folder should look like this
codetrain.scp  robotino.dfa   robotino.grammar  robotino.voca
prompts        robotino.dict  robotino.term
  • Now create a folder called 'lexicon' in the 'voxforge' directory.
cd ~
cd voxforge
mkdir lexicon
  • Download the voxforge_lexicon file and save it in the 'voxforge/lexicon' folder you just created.

Step 9: Record the training data

  • You must have a headset with a mic or a desktop boom mic. Preferably the same mic which will be used for speech recognition on the robot. Built in laptop or desktop mics are not recommended.
  • Create folder called 'train' in the 'voxforge/auto' folder and then a folder called 'wav' in the 'train' folder.
cd ~/voxforge
cd auto
mkdir train
cd train
mkdir wav
  • Open the prompts file from the /home/%yourusername%/voxforge/auto folder in a text editor (for example gedit).
  • Open Audacity and configure it as follows
    • In the Edit>Preferences>Devices (or Audio I/O); make sure that you select 'Channels: 1 (Mono)' under the 'Recording' section.
    • In the Edit>Preferences>Quality; make sure that the 'Default Sample Rate' is set to '16000 Hz' and the 'Default Sample Format' is set to '16-bit'
  • If you need to configure your USB Headset then do the following, otherwise skip this bullet point
  • Create a new text file called '.asoundrc' and open it in gedit as follows
gedit ~/.asoundrc 
  • Paste the following text in the file
 pcm.!default {
         type asym
         playback.pcm {
                 type plug
                 slave.pcm "hw:1,0"
         }
         capture.pcm {
                 type plug
                 slave.pcm "hw:1,0"
         } 
 }
  • Save the file and restart the computer.
  • Make sure your microphone volume in Audacity is set to 1.0.
  • Then click Record (i.e. the red circle button) and begin speaking in your normal voice for a few seconds, and then click Stop (i.e. the yellow square button). Look at the Waveform Display for the audio track you just created. The Vertical Ruler to the left of the Waveform Display provides you with a guide to your audio levels. Try to keep your recording levels between 0.5 and -0.5, averaging around 0.3 to -0.3. It is OK to have a few spikes go outside the 0.5 to -0.5 range, but avoid having any go beyond the 1.0 to -1.0 range, as this will generate distortion. If necessary, adjust Audacity's microphone volume to keep your audio within the proper ranges.

Audacity screenshot.png

  • To begin, you should not have any tracks displayed in the Audacity window. If you do, click the x icon at the top left of the audio track display (or hit ctrl-z as many times as is required to remove them; or restart Audacity). If you don't Audacity will happily record your new track, and leave your old track untouched, and when you export your audio to a wav file, both tracks will be merged to your wav file.
  • Make sure your volumes are set properly, as outlined in the preceding section.
  • Record you first file by clicking 'Record' in Audacity and saying the words in the first line of your prompts file:
ROBOTINO MOVE ROBOTINO ROTATE ROBOTINO STOP
  • Speak normally - not too slow or too fast - and clearly. Pause slightly before you begin speaking and leave a short pause after you have completed (i.e. half a second pause before and after you speak). Remember not to breath out until you have clicked stop - most microphoness pick up breathing noises.
  • Click the 'Stop' icon when you are completed.
  • Review your waveform to ensure that highest and lowest peaks of your recording is between 0.5 and 1.0 in the upper range and the lower range is between -0.5 and -1.0. If they are, then listen to the file (press 'Play' in Audacity) to make sure your pronunciation is clear and that you do not hear any non-speech noises (i.e. breathing noises, lip smacking, or background noises, ...). If there are any problems, hit ctrl-z and re-record your file.
  • If the file sounds OK then click File>Export and make sure that the format is WAV signed 16 bit PCM. Name the file as sample1 (for the first sentence) and save it in the 'train/wav' folder.
  • Repeat the same procedure for the rest of the sentences in the prompt file.

Step 10: Running the script

  • Create a new folder called 'scripts' in the 'voxforge/auto' folder
cd auto
mkdir scripts
  • Download and extract this file in the 'voxforge/auto/script' folder.
  • The scripts folder should look as follows
create_trainscp.pl  HTK_Compile_Model.sh  interim_files  perlsort.pl
fixfulllist.pl      input_files           logs
  • Now create a folder called 'mfcc' in the 'voxforge/auto/train' folder as follows
cd auto
cd train
mkdir mfcc
  • Now run the script 'HTK_Compile_Model.sh' from the 'voxforge/auto/scripts' as follows
cd voxforge/auto/scripts
./HTK_Compile_Model.sh
  • The script should create two file 'hmmdefs' and 'tiedlist' in the 'voxforge/auto/acoustic_model_files' folder.

Step 11: Copying the acoustic model files to robotino

  • Copy and replace the files 'hmmdefs' and 'tiedlist' in the '/etc/robotino/sr/julius/acoustic_model_files' folder with the ones you just created.
  • Robotino is now ready to recognize your speech input.