The idea of our project stemmed from seeing one of the previous ECE 4760 final projects, Musical Water Fountain. In their project , they used Fast Walsh Transform to analyze audio signal generated by a MP3 player (shown in table below).

Then they would turn on the LED that corresponded to the most energetic frequency division in the input frequency spectrum. This made us wonder if identifying speech is possibly by a method similar to this.

In fact, with today's technology, speech recognition is fully realizable and can even be fully synthesized. However, most of the software that deals with speech recognition require extensive computation and are very expensive. With the limited computation power of mega644 and a $75 project budget, we wanted to make a simple, smart voice recognition system that is capable of recognizing simple vowels.

After careful research and several discussions with Bruce, we found that vowels can be characterized by 3 distinct peaks in their frequency spectrum. This means if we perform a transform to input speech signal, the frequency spectrum profile will contain characteristic peaks that correspond to the most energetic frequency component. Then if we check to see if the 3 peaks in the input fall in the ranges we defined for a specific vowel, we will be able to deduce is that vowel component was present or not in the user's speech.

The main structure of our decoder system centers on the mega644 MCU. Our program allows the MCU to coordinate commands being placed by the user via PuTTY and the button panel while analyzing the user's audio input in real time. On the lowest design level (hardware), we have microphone and a button panel to convert physical inputs by the user into analog and digital signals the MCU can react to. On the highest level, PuTTY displays the operation status of the MCU and informs the MCU of user commands being placed at the command line. PuTTY also offers user the freedom to test the accuracy of our recognition and simulates a security system where the user must say a specific sequence of vowels to see a secret message.

Basically, the first three formant frequencies (refer to peaks in harmonic spectrum of a complex sound) can attribute to the different appeal of vowel sounds.

Therefore, if we can pick out formant by intensities in different frequency ranges, we can identify a vowel sound and use sequence of vowel to generate an audio pass code specific to that vowel.

The biggest difference between our analysis and musical intensity is that we need to adjust the frequency range stated above to better tell apart the difference between several peaks and combine all other information including amplitude. We need to decide which frequency transform algorithm is better to be used for a real-time audio addressing in both accuracy and computation speed. In fixed point DSP function in GCC, DCT, FFT & FWT are several common used algorithms. In our case, we chose Fast Walsh Transform over the rest simply because of its speed and its linear proportionality to Fast Fourier Transform.

The Fast Walsh Transform converts the input waveform into orthogonal square waves of different frequencies. Since we are only working with voice ranges here, we set the sample frequency to 7.8K which allows us to detect (ideally) up to 3.8kHz. We also knew that the lowest fundamental frequency of human voice is about 80-125Hz. Thus, we chose a sample size of 64 bit. This generates 32 frequency elements equally spaced from 0Hz to 3.8kHz (not including the DC component). The individual frequency division width is 3.8k/32=118.75Hz which gives maximizes our frequency division usage (since we could have useful information in every division instead of say a division width of 50Hz, where the first division does not provide useful information). Furthermore, this choice also minimizes our computation time since the more samples we have to compute, the more time it will take for the MCU to process input audio data.

In this part, most research we did were based on common vowel characters like 'A','E','I','O','U', which demonstrated that the method we attempt to develop could achieve. Yet in the real case, we found that the difference of these five characters is not as obvious as simply comparison between frequency sequency could distinguish.

We first use Adobe Audition to observe initial input waveform taken directly from Microphone and AT&T text2speech as shown in the picture. Although the waveform corresponding to the same vowel would result in a similar shape, there still exists difference which we may find more straightforward in frequency domain.

The first program in MATLAB is based on Prof. Land's code that compares the FFT and FWT outputs as spectrograms, then takes the maximum of each time sliced transform and compares these spectrograms. Top row is FFT power spectrum, FWT sequency spectrum is in the bottom. The maximum intensity coefficient of each spectrogram time slice in FFT and FWT are almost in the same shape. We'll take one spectrum as an example.

Another program directly implements FFT and show a frequency series. In this figure we can clearly see the resonance peaks of a vowel. This transform is 256 points. Also, notice that because of noise interference, it would be hard to tell apart the second peak for [EE] and this is not the only case.

Due to the limited precision of our Fast Fourier Transform, frequencies that differ by a value that is less than the width of our frequency division are often not distinguished. When dealing with boundary frequencies, this was a problem for us since the peak frequency did not always reside in the same frequency division. To improve upon this, we used multiple divisions but we still had errors since we cannot consider every possible boundary case. We improved upon this further by boosting the gain of our op amp from x10 to x100. This boast gave us a much better summary result and reduced our error. However occasionally, we still have errors that stem from the precision of our analysis tool.

The only standard applicable to our design is the RS232 serial communication protocol. We used a MAX233 level shifter and a custom RS232 PCB designed by Bruce Land.

The mathematical theories for frequency analysis of audio signals were obtained from both discussions with Bruce Land as well as R. Nave's webpage from Georgia State University.

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