Make your own free website on Tripod.com

An Experiment with EVPmaker and Synthesised Monosyllabic Words

A set of 13 random monosyllabic largely tri-phonemic words was prepared and typed into a Winspeech synthesiser on the PC. The 13 variations were intended to cover all possible vowel sounds.

The set of 13 words was then played through the synthesiser and the resulting Sound Card output fed to an AIWA HXPRO tape deck where the sounds were recorded.

On completion of the recording, it was played back into a Cool Edit system and a .wav file prepared. The section of that file containing the 13 words was highlighted and saved as Selection, the resultant file being then opened to enable final editing to get rid of empty passages at beginning and end.

The final result was a file of about 5 seconds long containing 13 words, thus giving each - on average, about 385 milliseconds.

The file was then loaded into EVPmaker and a number of conversion (EVP making) sessions were done using each using one from the following list of slice-size periods, 45, 90, 120, 450, and 900 milliseconds

Words and phrases did occur and were logged. At the end of each conversion session the result was examined in Cool Edit 2000 to pick up any words or phrases, and the duration of all the words or phrases were added up to give a total which was then compared with the duration of the original to give a "conversion efficiency".

Additionally the conversion session file was reversed to check for words in that mode. These did occur but were at a lower rate and were not logged at this time.

In normal mode there were a couple of cases of two voices saying different things simultaneously were observed and there was some question as to whether the duration for that example should be doubled, but at present dual voicing has been ignored.

The following table shows conversion efficiencies.

 

Duration of slice (ms)

Conversion Efficiency (%)

45

0

90

66.2

120

92.4

450

107

900

100

In the case of 450 there was a slight overlap between two adjoining segments which meant that the same phoneme(s) was used twice.

The score for 900 may be unrealistic as there was a flaw in the results. This was that two - or three of the words/phrases were repeated two times.

It came out along these lines,

'Now Time For, Now Time For, Ten Ten, Mao Ten, Hint look May, look May'

(total number of phonemes = 41)

I wont have to explain the significance of 10-10 - the day just past - to Bay Area readers. This is a big day of celebration for many Chinese, being the day on which Sun Yat Sen finally overthrew the last Imperial dynasty.

This could all be seen as a warning or as advice. But is it EVP?

Let me write down these words again,

1

2

3

4

5

6

7

8

9

10

11

12

13

Tan

ring

Ten

dan

Mouth

read

bug

Look

Make

time

now

Pore

sod

(Total number of phonemes 38)

From the above table, given that some words will be sliced through prior to completion, it looks as though the words selected by the random noise generator were,

10, 11, 12, 10, 11, 12, 3, 3, 5, 3, 2?, 8, 9, 8, 9

Well, that doesn't look very random to me. So either the RNG is not running right, or there is a paranormal effect.

The following standard measurements are suggested,

Conversion efficiency = total duration of all recognised words or phrases, divided by duration of original material, times 100, percent.

Gain - number of phonemes in the accepted words in the conversion, divided by the number of phonemes in original.

(In this case the gain would be given by 41/38 = 1.08)

At this time, when so much depends on judgement, they are probably only useful for relative measurements where only one analyst is involved.

Alexander MacRae