Despite the fact that we are processing and analyzing data in real time, the FWT analysis and summary were produced instantaneously after the release of the yellow button. There are no known errors associated with controlling the MCU operation via both PuTTY and the physical button panel. Even when 2 buttons are pressed at the same time, the system will sequentially execute the valid commands.

Here are some screen shots of our system during operation:

When the system is turned on, our MCU automatically enters the default testing mode. At the same time, PuTTY will display a welcome screen informing the user that the system is ready to take inputs.

In the testing mode, the user is able to see the summary results of only saying 1 vowel. Audio input is processed when the user holds down the yellow button while speaking into the microphone. When the yellow button is released, the program will automatically compute the FWT and display the prediction in PuTTY.

To leave testing mode, the user can just press and release the red button once. As shown below, PuTTY shows that the program has exited testing mode and entered the decoding mode.

In the decoding mode, the user can set a sequence of 5 vowels as the system password and repeat the same sequence via the microphone while holding down the yellow record button.

If the user accidently entered the reset command before setting a password, the system will inform the user that there is no valid password being stored at the time. The user should set the password first by entering 's' to the command line.

New password is entered by inputting the vowel sequence with commas separating each vowel input. Once entered, the system will display the entered result and automatically enter recording mode where the MCU simply waits for user's audio input.

If the input audio sequence agrees with the stored password, then the congratulations screen will appear along with the secret message.

Anytime there is a command prompt at PuTTY, the user can choose to reset his/her current audio input by entering 'r'. This erases all of the audio inputs stored so far in the system and allows the user to re-record the password again.

Anytime there is a command prompt at PuTTY, the user can see the stored system password by entering 'p' for print. The system will display the entered vowel sequence.

The user can reenter testing mode from decoding mode by entering 't' at the PuTTY command line.

Here are 2 videos demonstrating our system at work

We originally designed our program to decode female voices. However, when we tested our system, we discovered that it decodes male voices (of much lower fundamental frequency) just as accurate as it decodes female voices. However, due to the limited precision of the FWT we implemented, in cases where the frequency peaks are near our predefined characteristic peak value for a vowel, errors occasionally occur. We tested our program with a couple of our friends and for a male voice, the program is able to accurately predict the vowel said 49/50 times and for female voices, the program is able to accurately predict 45/50 times. Furthermore, the program only accurately recognizes vowel is the user is consistent in speaking (no accents or instability during recording).

We tested our program with a couple of our friends and for a male voice, the program is able to accurately predict the vowel said 49/50 times and for female voices, the program is able to accurately predict 45/50 times. Furthermore, the program only accurately recognizes vowel is the user is consistent in speaking (no accents or instability during recording).

We also found that the MCU tend to confuse between "OO" and "EE" or "OH" and "AE". In the case of "OO" and "EE", the waveforms are very similar. In the FWT output, both vowels have peaks that often overlap. In our program, "OO" and "EE" are determined by the maximum amplitude obtained in the transform. In normal speech, "OO" is louder and "EE" (see below for waveform comparison). This explains why MCU mistakes one for the other.

In the case of "OH" and "AE", FWT of the input waveforms produce almost the same first and second peaks. The two vowels are distinguished mainly based on the location of the third peak. However, the amplitude of the third peak is relatively low and can be easily mixed up with noise. Thus, predictions made about "AE" and "OH" and differ greatly depending how the speech was formulated.

Here are some test results we got using our system:

The system that we have designed can be used as a basis for implementing speech recognition since speech consists of vowels and consonants that can be identified using frequency analysis. An example of possible implementation in the real world would be using speech recognition in security systems, something that could be more convenient than entering passwords on a keypad to people with less proficient vision.

Furthermore, our system is simple and easily to handle. The only precaution in using our prototype system is the user must be careful in touching the PCB and port pins to minimize ESD hits.

Copyright ECE 4760 Spring 2011 Annie (Wei) Dai and Youchun Zhang