Pocketsphinx with default model has zero accuracy with my accent

[Dmitry Pryadkin]

I'm new to sphinx so please excuse me for this question. I'm not a native English speaker either. But pocketsphinx does not understand just anything I say to it.

I've been reading through the FAQ section, I'm sure that my recordings were of the right format, that is 16khz, signed little endian, 16 bit.

I'm creating a Unity plugin which I'm going to publish on Github, this is how samples are taken:

        float sample = samplesBuffer[i]; // [-1; 1]
        short sample16Bit = (short)(sample * (65535 / 2));

I also tried to record my voice with Audacity, the accuracy is still zero.

I used pocketsphinx_continuous.exe, it correctly recognizes goforward.raw from pocketsphinx, but it does not recognize my own version of goforward.raw from https://github.com/Bangybug/blog/blob/master/question-pocketsphinx/goforward.raw

The log is here:

pocketsphinx_continuous.exe -infile goforward.raw -hmm c:\workspaces\sphinx\pocketsphinx\model\en-us\en-us -lm c:\workspaces\sphinx\pocketsphinx\model\en-us\en-us.lm.bin -dict c:\workspaces\sphinx\pocketsphinx\model\en-us\cmudict-en-us.dict

https://github.com/Bangybug/blog/blob/master/question-pocketsphinx/log.txt

Do I have to train it for my own voice/recording/accent?

[Nickolay V. Shmyrev]

I'm not a native English speaker either.

In the future avoid this kind of entry, your English is good and it does not bring any new information. Respect yourself.

Do I have to train it for my own voice/recording/accent?

This problem has nothing to do with the accent. Your audio is too loud and short and recognizer can not adapt to the volume level quickly, initial CMN estimation is wrong. Second utterance after first one must be decoded properly if you keep CMN state. You can read discussion here for more details:

https://github.com/watsonbox/pocketsphinx-ruby/issues/10