Various VSA software programs offer a graphic representation derived from that obtained on the Polygraph.
Alpes Software uses sound recordings made in WAV format, to 11,025 Hz, 8bits. The sound is filtered by retaining low frequencies. The envelope is not the most interesting and the modulation is added in order to bring out the low frequency oscillations.
Some VSA software programs use a theory introduced by the British physiologist Olaf Lippold, 1957, by extracting the low frequency of about 8 to 12 Hz and deducing a value of stress.
Understandably, it's not too difficult to compute the frequency for the curve displayed in Image 1, but it becomes much more complex for a curve as displayed in image 2.
Alpes Software, which is the basis and the software designer of Diogenes Lantern Pro now uses a new technique for calculating the value of stress, and not based on a direct calculation of the micro-tremor
Alpes Software uses a knowledge base collected from several VSA experts. Thousands of samples were processed, with approximately 1,200 being retained.
A totally independent scoring excercise was conducted by VSA experts. At the culmination of the project it was notable that 80% of the results presented only 5% difference on the estimated values of stress, which provides a consistent and reliable basis for the techniques used by Alpes Software.
The technique consists in training a neural network with 1000 phonemes (utterances) from interrogations. The voice samples are processed to obtain the curves as per Image 1 and Image 2.
These graphs were digitized to provide a basic set for training the neural network. The training of the neural network generates a file of weights that are at the heart of the network processor. Some parameters are used to adjust the neural network ; ie : the number of neurons in input, the threshold, the desired error (1%). The duration of training is approximately 45 minutes on a standard PC.
A fundamental property of the neural network is it’s capability to produce a reliable result when we enter a voice graph for which it has not been trained. When we submit a graph for which it was trained, it will produce a result with the specified error (1%).
The software AVSA extracts in real time phonemes, digitizes curves, charts are submitted to the network, then displays the results.
Another interesting neural network property is that there is no need to formalize mathematically complicated formulas, to extract frequencies of micro-tremors. The neural processor teaches itself how to deduct the outcome on the basis of data provided.
It is totally free from the assumption 12Hz for low stress and <8HZ for the presence of stress. Our only certainty is the know-how of human experts who have conducted thousands of successful vsa examinations, together with known solution outcomes.
The AVSA software can calculate up to 20 estimates of stress per second on a standard PC. Chart tests (validated protocols) are scored within 2 seconds. Exam scores (average of the tests conducted) are displayed instantly at the culmination of the last test chart.
AVSA PRO currently provides TWO commercial versions :
Version 2,0AG : For structured DOD interview examinations + all auto features. Manual operation is still retained.
Version 1,7KT : For free flow, real time narrative screening. Easy to follow graphics for operators processing high volumes of subjects.
Two stress levels engines are provided.
Both programs feature Autogain, to compensate for low input volumes. AVSAPRO is fully compatible with all versions of Microsoft Windows OS.
Both programs are provided per 1 licence purchased.