5 - Noise triggered impulses for EVP voice generation - by Andrés Ramos

[Andres Ramos]

1.  Abstract

It was more by accident that we experimented with noise impulses in our setup because they were so abundant in the signals we recorded.  First we thought that those spikes were just interference signals from common electric sources like power supplies, electrical machine, CFL a.s.o. After I denoised a sequence of spikes by accident I was amazed to find out that they also contained speech patterns. That was the start of a quest for the meaning and characteristic of spirit spikes or spirit impulses as we called them since we knew they were not interference. The fascinating thing was that by proper signal post processing we could recover a good voice part even from just a small bunch of impulses. It was logical to devise a test setup that was designed expecially to scrutinize these impulses we so far just analyzed as a byproduct of other experiments. I wanted to find out if we could gain good quality voices from pure impulses because this digital representation as impulses is practically noise free.

 

2. Test setup

The challenge was to design a circuit that uses a noise signal as an input signal and generates random impulses from them. I had the idea of making two independent noise sources with the very reliable OA9 diode. The preamplified noise would trigger the set- and reset inputs of a simple RS-Flipflop. The idea was that randomly setting and resetting a flipflop would generate trains of impulses with varying frequency and duty cycle.

Electronic schematic of first RS-FlipFlop EVP receiver

The schematic shows very clearly the two indentically designed noise generators. The trigger sensitivity for each R- or S-channel can be adjusted independently. I provided two LED's with different colours to indicate the SET or RESET state of the FlipFlop. The idea was to give the operator an indication of the triggering count and if the triggering is symetrically or not. For this test I made a breadboard design and no real prototype.

 

3. Test results

The circuit worked pretty much as expected and the signal was very symetrically. That means that SET and RESET events were more or less in equilibrium.

The results were very encouraging if a reasonable amount of denoising and hp-filtering in audacity was applied. I was amazed since it was proven that voice patterns coded in impulses can be recovered up to a certain amount that makes the voice intelligible again.

You can hear a collection of samples here.

 

4. A new design

The results of the first design were good but I valued it as a bit too complicated. My consideration was that maybe one noise source could do as well and I wanted to use the excellent trigger facilities of the famous monostable multivibrator circuit NE555. That gave me the base of a new design I made 9 months after the first one

Electronic schematic of 2nd design

You see that this schematic is much simpler. It contains roughly only half of the components from the first design. Here is only one noise source. The amplified signal is feeding the trigger and threshold inputs of the NE555. Basically the trigger input corresponds to the SET input of a RS-FlipFlop and the threshold can be seen as a reset. The NE-555 triggers its internal flipflop  if the voltage on Pin 2 undergoes a trigger level of UB/3 = 4V. The output (Pin 3) is set to HIGH then. If the voltage rises above 2*UB/3 = 8V the circuit resets the internal flipflop again..

 

5. Results of test with 2nd design

Obviously the distribution of impulses is not that symetrical as with the first design but this had no remarkarble influence on the signal quality. The first track shows the raw signal. The second track shows the signal after processing with Paulstretch. This is an intelligent algorithm in audacity designed to stretch a recording by filling the gaps that naturaally occurr by stretching with data that was synthesized after the specification of the raw signal. In our practice Paulstretch proved as very valuable to convert impulses in readable voices. The third track is the signal from track 1 after 18 dB of denoising which is moderate.

The picture above shows the impulse trains in zoomed presentation. As expected frequency and duty cycle are varying randomly. A sequence of experiments showed that much impulses do not necessarily lead to good voices despite as one might think because more impulses mean more entropy and more opportunities for the spirits to form voices. This is not the case. In fact it proved that less impulses give better results than much impulses. The reason is that less impulses are an outcome of higher SET/RESET levels on the NE555. Only the amplitude maxima (positive and negative) are triggering impulses thus the noise below does not trigger impulses and only the high impulses of strong voices are making it through. See it as a special form of signal to noise improvement in the time domain. The picture below shows a signal with less impulses in three parts. Left is the raw signal, in the middle the signal after Paulstretch and on the right after denoising

 

Hear the audio sample corresponding to the picture here. You can hear that the rhythm of the speech is very good perceivable even in the raw signal. After paulstretching the result is of so good quality that denoising is almost obsolete.

The last picture shows the spectrum of the raw signal. The spiky outline already shows a good modulation with voices. You can hear samples made with this design in the same audio directory as specified above.

 

 

6. Conclusions

It was proven that impulses triggered by a random noise signal are representing a spirit voice signal in an, I am tempted to say, digital form with very good signal to noise ratio. The paulstretch function in audacity is an excellent tool to convert the impulse trains into a readable signal.