speech model basically consists of two parts:

• Language model: Describes all existing words and what sentences are grammatically correct. It will be created from your active scenario
• Acoustic model: Describes how words sound. It will be either built solely through your voice recordings (training) or with the help of a base model

SPHINX backend“ uses:

• PocketSphinx recognizer
• SphinxTrain model generation backend

models:

• US English HUB4 WSJ SPHINX model. This model is part of pocketsphinx-extra where it is called \”hub4wsjsc3s8k.cdsemi5000\“.

prereq

sudo apt install bison gradle openjdk-8-jdk

build from sources http://cmusphinx.sourceforge.net/wiki/download, not use ubuntu packages

sudo apt install simon
 
sudo apt install libsphinxbase3 python-sphinxbase sphinxbase-utils pocketsphinx sphinxtrain

edit

/usr/lib/sphinxtrain/etc/sphinx_train.cfg

$CFG_CD_TRAIN = 'no';

sudo apt-get build-dep simon
 
git clone https://github.com/KDE/simon.git
cd simon
export QT_VERSION=qt4
rm -fR build
./build_ubuntu.sh

http://cmusphinx.sourceforge.net/wiki/tutorialadapt

sudo apt install pocketsphinx-en-us/xenial

training from http://cmusphinx.sourceforge.net/wiki/tutorialamdown

• download http://www.speech.cs.cmu.edu/databases/an4/an4_sphere.tar.gz
• enter and sphinxtrain -t an4 setup

For small amount of data for training without context-dependent models use continuous model instead of semi-continuous one and this error will be solved.

This warning is caused by the fact you disabled context-dependent training and used semi-continuous model. You can ignore this error until you have more data and use context-dependent training.

training

mkdir /tmp/test
cd /tmp/test
 
sphinxtrain -t test setup

now we have

.
└── etc
│   ├── feat.params
│   ├── sphinx_train.cfg
│   ├── test.dic
│   └── test.filler

adjust sphinx_train.cfg TODO