Swig
brew install swig

$ bash download.sh

Build and install SphinxBase

$ cd sphinxbase  
$ ./autogen.sh  
$ ./configure  
$ make  
$ sudo make install  

Export the environment variables to the ~/.bash_profile:

export LD_LIBRARY_PATH=/usr/local/lib
export PKG_CONFIG_PATH=/usr/local/lib/pkgconfig

Build and install PocketSphinx
$ cd ../pocketspinx
$ ./autogen.sh
$ ./configure
$ make
$ sudo make install

Build and install Sphinxtrain
$ cd ../sphinxtrain
$ ./autogen.sh
$ ./configure
$ make
$ sudo make install

Create a directory $ mkdir kws $ cd kws

Move the mandarin acoustic model to the current work directory
$ mv ../zh-cn .

And then unarchive the file. Download the Mandarin dictionary from

https://sourceforge.net/projects/cmusphinx/files/Acoustic%20and%20Language%20Models/Mandarin/zh_broadcastnews_utf8.dic/download

Here are the files you will need:

• corpus.txt
• corpus.lm
• corpus.dic
• recodings of you speaking each keyword
• corpus.fileids
• corpus.transcription

Write down the keywords in corpus.txt, one keyword in a row respectively. For example:

啟動
暫停
關閉
恢復
洞見未來
KKBOX

Then go to the LMTool page. Click "選擇檔案" to upload corpus.txt and click "COMPILE KNOWLEDGE PAGE". Download the file followed by the extension .lm and .dic, and then change theses two files name into corpus.lm and corpus.dic respectively.

According to the previous example, you need 5 recordings and each of them records one keyword that you speak. Each recording file should be followed by the extension .wav recorded at the sample rate of 16k Hz in mono with a single channel.

Supposed that the file name of recordings are:

voice_0001.wav
voice_0002.wav
voice_0003.wav
voice_0004.wav
voice_0005.wav
voice_0006.wav

You need to create a file called corpus.fileids and write down:

voice_0001
voice_0002
voice_0003
voice_0004
voice_0005
voice_0006

Create a file called corpus.transcription, write down the content of each recording row by row, and specify the file name at the end of each row. For example:

<s> 啟動 </s> (voice_0001)
<s> 關閉 </s> (voice_0002)
<s> 暫停 </s> (voice_0003)
<s> 恢復 </s> (voice_0004)
<s> 洞見未來 </s> (voice_0005)
<s> KKBOX </s> (voice_0006)

Make sure that, before and after the content there are an <s> an </s>, and then specify the file name of recording in the form of (recording file name).

Open the corpus.dic and the zh_broadcastnews_utf8.dic which you just downloaded. First, delete the phonemes after KKBOX, because we want to map KKBOX to phoneme from Mandarin dcitionary. Find the phoneme mapping each mandarin word you need in zh_broadcastnews_utf8.dic. For example, you need to find the phonemes mapping the following words: "K", "BOX", "啟", "動", "關", "閉", "暫", "停", "恢", "復", "洞", "見", "未" and "來". After finding out the corrsponding phonems, copy and paste to the corpus.dic. Then you should have the corpus.dic look like:

KKBOX k ei k ei b a k e s
啟動 q i d ong
恢復 h ui f u
暫停 z an t ing
洞見未來 d ong x ian w ei l ai
關閉 g uan b i

Besides, don't forget to put space between each phoneme.

Now, the file structure should look like:

voice_0001.wav
voice_0002.wav
.....
corpus.fildids
corpus.transcription
corpus.lm
corpus.dic
corpus.txt
zh-cn
zh_broadcastnews_utf8.dic

sphinx_fe -argfile zh-cn/feat.params \
    -samprate 16000 -c corpus.fileids \
    -di . -do . -ei wav -eo mfc -mswav yes

You should have the following files in yours working directory:

voice_0001.wav
voice_0001.mfc
voice_0002.wav
voice_0002.wav
.....
corpus.fildids
corpus.transcription
corpus.lm
corpus.dic
corpus.txt
zh_broadcastnews_utf8.dic

pocketsphinx_mdef_convert -text zh-cn/mdef zh-cn/mdef.txt

cp /usr/local/libexec/sphinxtrain/bw .

cp /usr/local/libexec/sphinxtrain/map_adapt .

cp /usr/local/libexec/sphinxtrain/mllr_solve .

./bw \
-hmmdir zh-cn \
-moddeffn zh-cn/mdef.txt \
-ts2cbfn .ptm. \
-feat 1s_c_d_dd \
-svspec 0-12/13-25/26-38 \
-cmn current \
-agc none \
-dictfn corpus.dic \
-ctlfn corpus.fileids \
-lsnfn corpus.transcription \
-accumdir .

./mllr_solve \
-meanfn zh-cn/means \
-varfn zh-cn/variances \
-outmllrfn mllr_matrix -accumdir .

You now need to copy the acoustic model directory and overwrite the newly created directory with the adapted model files:

cp -a zh-cn zh-cn-adapt

And then,

./map_adapt \
-moddeffn zh-cn/mdef.txt \
-ts2cbfn .ptm. \
-meanfn zh-cn/means \
-varfn zh-cn/variances \
-mixwfn zh-cn/mixture_weights \
-tmatfn zh-cn/transition_matrices \
-accumdir . \
-mapmeanfn zh-cn-adapt/means \
-mapvarfn zh-cn-adapt/variances \
-mapmixwfn zh-cn-adapt/mixture_weights \
-maptmatfn zh-cn-adapt/transition_matrices

pocketsphinx_continuous -inmic yes -hmm zh-cn-adapt -lm corpus.lm -dict corpus.dic -mllr mllr_matrix

Done!

--
Reference: CMUSphinx