Strategy of implemantation.

[Cornebidouil]

Hi, I would like to try to make a speech recognition of short magic formula, so I would know your opinion on the way of doing it.
In my mind I need to create my own acoustic model from my recordings, with this in mind I have some question :
I have a recording file of all my formula but I cut it into records of one to three words (for each formula) so it last below one to two seconds. Do I need to regroup these recordings (to have files above 5 seconds) ?
For the testing part I had in mind to use my big recording files to do it, is there a length limit ?

Edit : I have forgotten to add that I use french phonems because my formula are close to latin, so maybe it's better to adapt a french acoustic model ?

Thanks in advance !

[Nickolay V. Shmyrev]

I would learn about concepts of keyword spotting first and build a test set to estimate the current alarm rate and precision for your phrase.

[Cornebidouil]

Thanks for your reply,
I am creating my keyword list so I need to estimate the threshold, how can I distinguish false alarm and missed detections ? (I need to find the limit of unique detection correct ?)
Second question what length of wav file in need to perform this determination ?

Example of my results :

pocketsphinx_continuous.exe -infile model1\fr-FR\allsorts.wav -hmm model1\fr-FR\fr-FR -dict  model1\fr-FR\spells2.dic -keyphrase Accio -kws_threshold 1e-20 -time yes -logfn my_log.log -beam 1e-200

Accio
Accio 1.510 1.730 0.958290

Does it mean that have just a positive result ? and this result is find at 1.51s and end a 1.73s for a length of 0.958290 ?

Sorry for these beginner questions.

[Nickolay V. Shmyrev]

Second question what length of wav file in need to perform this determination ?

Tutorial https://cmusphinx.github.io/wiki/tutoriallm/#keyword-lists says

Take a long recording with few occurrences of your keywords and some other sounds. You can take a movie sound or something else. The length of the audio should be approximately 1 hour

Does it mean that have just a positive result ?

Yes

[Cornebidouil]

Thanks, in case of it's words that I have invented I need to record myself during 1 hour ? Even if the objective is to detect the word on a sample of few seconds ?

[Nickolay V. Shmyrev]

Looks like you have trouble to read. I already quoted you above You can take a movie sound or something else.

[Cornebidouil]

Holyshit it's late sorry ...