- Speech Recognition Device
A Few Words About My Project ...
Hej ! Jag heter Rémy och jag är fransman.
Jag kommer från Paris där jag studerar elektroteknik på
Ecole FRancaise d´Electronique et d'Informatique
But so far, I am so lazy that I will carry on in english ...
Because I am fond of signal processing, I have tried to build a 68008 micoprocessor
based device which will recognize different speakers. I did not intend to create a
revolutionnary design but just to put my signal processing knowledge into practice.
The main originality of my work is that I do not use any D.S.P. chip like
the famous TMS320xx serie but this very simple Motorola Microprocessor.
So far I could not perform any spectral analysis
( FFT like or whatever ... ) which are very very veeeeeeeeery long to proceed
whith such a slow processor. I finaly switch on an other method known as
AUTOREGRESSIVE MODELLING that consists of creating
a mathematical model that reflects some of the statistical properties of the signal.
( It may seems awfull but it's not .... )
Althought, as far as I was alone to deal with the entire project, I have chosen not to
build a very complex electronic part which could have failed, but just an efficient one
that allow me to sample an analogical signal (10 Khz slow but sufficient ), to store it
in R.A.M. and to proceed some shrewd algorithms which lead my speech recognition.
So, I have been able to spend more time on the software part which has recquired
quite a lot of time because of the piece of research I had to carry out.
The main electronical architecture is very classical :
- 1 68008 the Motorola microprocessor
- RAM, EPROM ( to store the program ... ) and E2PROM to store the reference samples
- 2 PAL that perform the adress switching.
- 1 anti-aliasing filter ( didn't I speak about sampling ???? )
- Some other basical features such as 74HC373 for I/O communication ( Keyboard, leds )
or operational amplifier for the analogical part.
- 1 microphone (it can help ... ).
For those interested by the theorical aspect of my work : ( yes, some are ... )
I use the very famous Auto-Regressive model with 10 coefficients ( AR(10) ) to
proceed my recognition. This model reflects the physiological configuration
of the throat and is use in a lot of speech applications ( like Linear Predictive Compression ).
Althought, I compute the autoregression vector by solving the Yule-Walker equations
thanks to the Levisson Durbin Algorithm. I obtain a set of AR(10) coefficients that
are different for each person.
The project finaly worked and performed a very good analysis over a 500 samples set
and proved that it is possible to handle signal Processing application with cheap chips !