ASSAP: Paranormal Research
ASSAP: Paranormal Education
Privacy and cookie information ASSAP mailing list
 
 

An 'EVP Gallery' with a difference! You will not be told what to expect before you hear each sound. Judge for yourself and then compare it with other answers submitted already. These are more experiments or demonstrations rather than just plain EVP sound clips.

Before you start: Since there are discussions of these clips' contents lower down this page, don't get ahead of the text. After listening to each sample, follow the link immediately after it to get to the discussion section and then use the 'Return to experiment' link to return to your previous position. You will get more out of the experiments if you do each in turn and don't skip ahead! Most computers will have no problem playing these clips. If you DO have problems, look at 'Advice about playing the clips' below. Now go to the first experiment by following this link.

Advice about playing the clips: The files are in WAV format. You may need to associate WAV files with a default sound application on your computer for the links to work. You shouldn't need to use headphones to listen to these clips (indeed it may be inadvisable as some are loud!). Ordinary computer loudspeakers should work fine though you may need to have the volume set high for some of the clips.

Related topics: You will not find advice on recording EVP ('electronic voice phenomenon') and paranormal sounds here but try this page instead. To find out more about analysing EVP and paranormal sounds, read this page.

Experiment 1: Sound editing

Listen to sample 1 - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link. If you have any problems playing the sound, see 'Advice about playing the clips' just above.

Now look at the spectrogram of this sample:

Sample 1 spectrogram

Frequency increases downwards in this spectrogram with time increasing rightwards. The warmer colours (reds and yellows) are higher volume and the cooler ones (blues and greens) lower. Note there is no sound at all at the bottom. That's because all sound above 2000 Hz has been filtered out. There are some formant-like peaks (yellow) at around 1500 Hz which no doubt lead to the impression of a voice. Note, however, that all other formant frequencies are missing. Formants are the frequencies used by the brain to distinguish the individual sounds (phonemes) that form human speech.

Now listen to sample 2 - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Look at the spectrogram of this sample:

Sample 2 spectrogram

This is essentially the same except that the frequencies above 2000 Hz have been restored. With all these extra frequencies, the sound is naturally less distinct and noisy.

Now listen to sample 3 (loud compared with previous samples) - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Looking at the spectrogram:

Sample 3 spectrogram

This looks similar to the previous spectrogram except that there is less variation. Note that there is a lot more dark blue. This is the original recording. It is white noise produced by a radio tuned between stations.

Now listen to sample 3c - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

How the clips were produced

Spoiler warning! You should do the experiment before reading this section! So where has the reasonably obvious 'voice' that we hear in sample 1 (and maybe sample 2) come from? It actually started with sample 3, the white noise.

To get from sample 3 to sample 2, the recording was passed through a noise removal filter. This had the effect of removing or reducing some of the frequencies present. The noise removal process tends to accentuate the largest peaks (and the lowest troughs between them) in the frequency spectrum of the sound while removing the smaller variations. You can see this in the spectrogram. The noise filter was used at its lightest setting so that it did not cause huge changes. Nevertheless, it made a big difference to what could be heard.

To get from sample 2 to 1, all the frequencies above 2000 Hz were completely removed. This left the recording sounding distinctly voice-like, particularly if played repeatedly.

Sample 3c was produced by reversing sample 1! Some people claim they can hear EVP messages when the recordings are played backwards.

So the question remains. How did words apparently come out of random noise? And how can such radically different interpretations occur on different hearings (see sample 1 notes)?

By applying noise removal, the frequency peaks were accentuated making them more easily heard. The same process applies not only to sound frequency but also to time. So there are distinct time gaps formed between previously connected peaks. This gave the impression of more interrupted sound, more like a voice, rather than the original continuous hiss. Removing higher frequencies reduces distracting detail. In telephones, the audio range is restricted to 3000 Hz and voices can still be heard perfectly well.

By exaggerating the formant frequencies and any gaps between them, it was easier to make the brain think it is hearing words. The extraordinary variation in interpretations is due to the fact that there is insufficient information for the brain to know where 'words' start and end and where any gaps are. Using verbal transformation, your brain tries various likely interpretations which vary from go to go. It may even be influenced by background noise at the time you are listening to the recording.

The point of this experiment was to see whether, starting with white noise, it is possible to produce a reasonably convincing 'voice' simply from processing the sound through software. Radio white noise was chosen because it contains more variations than pure electronic white noise. The software was able to build on the slight variations in the radio white noise. The white noise sample used was deliberately chosen because it did not contain any obvious voices. Here is a longer sample 4 (loud and quite large compared with previous samples) of the same original radio white noise containing our chosen slice. Can you pick out which bit was used for the earlier samples? Can you hear any obvious voices? Try listening to sample 4 and then sample 3 again. Can you pick out sample 3 in sample 4 now? Does sample 3 sound the same as before? Then try listening to sample 1 again for comparison. Click this link for some observations on sample 4.

Discussion of Experiment 1

Spoiler warning! You should do the experiment before reading this section! Experiment 1 set out to see if, simply by applying software noise reduction and filtering, it is possible to turn a randomly selected sample of noise (with no obvious voices in it) into a 'voice' with apparently recognisable words. People listening to the samples will no doubt disagree on the results. This is to be expected given the subjective nature of hearing. Some will hear 'voices' in some or all the samples in this experiment, others will hear nothing but noise in any of them! People who do hear 'voices' will disagree about what they say and may be influenced in the interpretations given here, by what others report and by the order in which the clips are played. If the clips are played in a continuous loop, many people will hear the 'word' content apparently change over time (verbal transformation effect). They may also hear different 'words' on different occasions when they play the clips.

So far as can be ascertained, the original radio white noise sample used does not contain any obvious voices. The frequency spectrum is fairly random with minor peaks around 1500 and 3000 Hz. Noise removal software tends to accentuate the larger variations in the sound and remove the minor ones. If there is a real voice in a noisy background, noise removal software will make it clearer. However, the software is not intelligent. If it is given a sample of sound containing random noise and no real voice, the result can still be voice-like if there happen to be chance frequency peaks in original sound. You can see in the spectrograms in Experiment 1 how certain frequencies have been accentuated by the process.

The results of this processing sounded like anything from noise to crystal clear voices. The extraordinary range of message content interpretations, including contradictory sound sequences, backed the idea that the words were illusory (assuming you heard them at all!). The phoneme restoration effect helped to produce the variations by providing 'missing' phonetic sounds. Also, did you notice notice the highly unnatural rhythm (syllables speeding up and slowing down unexpectedly) and inflexions (intonation rising and falling in unexpected places) in the apparent phrases? Some people will have noticed that word interpretation could change depending on what sequence you played the sound clips in and if you were reading one of the written interpretations. All of this points to the crucial importance of context (with different length clips apparently affecting each other) and suggestion (where written interpretations changed what was heard). It shows that human hearing is largely in the brain, rather than the ear, and the system can be fooled by priming and expectation. The various anomalies in the 'voice' were caused by the brain attempting to interpret ambiguous sounds as speech.

Intonation

The changes in intonation were particularly interesting. In essence, your brain understands speech by recognising certain ratios between sound frequency peaks. In the case of voiced vowels, the relationship is harmonic (whole number multiples of the fundamental frequency). These harmonic relationships, in turn, imply the existence of a fundamental frequency which is normally the one provided by 'voicing' (though the fundamental frequency is only normally present in certain phonemes and is not necessary for a sound to be understood as speech). As this fundamental frequency changes it alters the perceived intonation of a voice.

The unexpected changes in intonation in this experiment probably occurred because the harmonic relationships appeared to change with time through the sample (in a way that would not happen in real speech) and so did the implied fundamental frequency. This is to be expected when the apparent harmonic relationships are just coincidental rather than from real speech.

Try this sample. It is the same as sample 1 but with all sound below 700 Hz removed. The chances are, you can't hear voices any more, even though most of the sound is still present, including the important peak around 1500 Hz. The harmonic relationships, spurious in the first place, have been broken by the removal of those lower frequencies and the 'voice' has gone. If you remove these same frequencies from a recording of a real, natural voice it become distorted but you'll still recognise the words. Real speech is robust enough to survive such treatment while our faux voice cannot.

Spectrum envelope

So why does sample 1, with every frequency above 2000 Hz removed, sound more 'voice-like' than sample 2, which has only had the noise removed? Scientists have discovered that it is possible for people to understand speech when it is artificially 'moved' to a new frequency range. This means that all the formant frequencies retain their characteristic ratios to each other but they are all shifted higher or lower than normal speech frequencies. Shifting the frequency lower leaves speech easier to understand than shifting it higher. Hence, it is more likely that humans will hear apparent 'voices' in noise at lower frequencies than at higher ones. So, removing the higher frequencies may make the apparent voices appear clearer. In addition, voices are easier to understand if they remain within a normal 'spectrum envelope'. A spectrum envelope is, roughly, the overall range of frequencies making up the complete voice sound (see diagram below). This 'envelope' gives the voice its intonation, allowing us to say if it is male or female, child or adult, etc. So, if the formant frequencies are shifted BUT kept within an expected spectrum envelope, the voice is easier to understand. It thus probably makes 'sense' to the brain as a single voice. So, in sample 1, restricting the frequency range to an envelope of 2000 Hz may have contributed to it sounding more like a voice. In sample 2 the frequencies are unrestricted which may make it sound more like the noise it really is and less like a voice.

Spectrum envelope

As a further experiment here is another sample to try. Try repeating this one and see if you can hear any voices and, if so, what are they saying? It will sound like a voice to some people (particularly if they heard the 'voice' in sample 1) but probably more difficult to make out. Maybe it's saying 'tell us your secret' or 'queen of Sheba' or something else entirely. Try comparing it with sample 1.

This new clip was produced from sample 2 (the 'noise removed' version) like sample 1. However, instead of restricting the frequencies to 0 - 2000 Hz (as in sample 1), all the frequencies EXCEPT 500 to 2500 Hz were removed. So, the total frequency range was the same except that everything was shifted higher.

In the diagram (left), frequency (bottom scale) is shown against sound intensity (left scale) for the latest sound sample. Notice how the sound is restricted to the area between 500 and 2500 Hz. The overall outline of the area in red is the 'spectrum envelope' for this sample (a more precise definition of the term 'spectrum envelope').

If you heard the voice in the latest sample you probably noticed several differences between it and sample 1:

  • the new sample 'voice' sounded higher in tone
  • it was more difficult to make out the 'words'
  • the rhythm, though similar, was less pronounced and slightly different
  • the 'words' (if you heard them) were probably different

The frequency shift changed the tone of the voice upwards, as you'd expect, making it sound like a different 'person'. The 'words' were different and more difficult to make out because there were fewer and different chance frequency peak ratios present in the noise. Also, the different pattern of peaks would have altered the apparent word spacing.

If you shift the spectrum envelope even higher, to say 1000 to 3000 Hz, the 'voice' disappears completely and the result sounds like noise. This is probably because of a complete lack of strong chance frequency peak ratios in that area.

Rhythm

The 'unnatural rhythm' of the 'voice' in this experiment would have been caused by the random spacing in 'consonants' and 'word' gaps. Because these were not real, the brain did its best to interpret the 'words' by altering the perceived timing. Sometimes you can hear alternative 'word' layouts on repeat listenings to sample 1. Yet another sign that this was not real speech.

Another thing you might have noticed is that, if you could not understand sample 1 when you first heard it, the 'words' may have 'appeared' spontaneously in your head a few seconds afterwards. This is your brain reviewing your 'acoustic memory' (a type of short term memory lasting seconds), even when you were no longer hearing the sound, and doing a lexical analysis. This delay effect is more noticeable when words are difficult to hear (such as in a noisy environment) or, as in this case, don't really exist at all.

Message content

Scientists have found that there is strong 'top-down' influence involved in word recognition (the 'interactive activation model'). Thus, words are imposed on phonemes by the brain, even if they don't quite fit, to make sense of speech. This explains the verbal transformation and phoneme restoration effects. It also explains why you are likely to hear complete words, rather than random sounds, in voice-like meaningless noise. If the process were 'bottom-up' (instead of 'top-down'), it would be more influenced by the lowest levels sensory inputs ie. the sounds themselves. Our method of hearing allows us to hear speech well in noisy surroundings, even if some of what we think we 'hear' is actually unconscious guesswork based mainly on context. Those people who could not hear any words at all, in the samples, were hearing the noise as it really is.

Reverse messages

Sample 3c was sample 1 played backwards, something that is easy to do with audio editing software. The phenomenon of 'reverse message' EVP apparently arose because sometimes a recording may sound like a voice but makes no sense. Playing it in reverse can, apparently, sometimes turn it into an intelligible message.

In this experiment, there were apparently intelligible 'messages' when played in either direction, though the content was different. Sample 3c didn't sound like the recording of an ordinary voice played backwards. That's because, of course, it wasn't - it was just processed noise. Noise it still noise no matter what direction it is played in. There are still apparent formants in it but they are just in a different order, hence the different messages. There is no reason at all why voice-like noise shouldn't work in any direction.

Reverse messages may be seen as a sign of the paranormal, since they would never occur in nature. However, once again, voice-like sounds can be played just as well backwards. So such messages cannot be taken as a sign of the paranormal at work on their own.

Conclusions

Crucially, this experiment demonstrates how even light use of audio editing software can turn more or less random noise into something voice-like. This is one of the reasons why it is advisable not to edit your recordings at all. Also, note how written interpretations can affect what people subsequently hear! Never tell people what to expect when you give them an EVP, or other paranormal sound, recording to judge.

From a wider perspective, this experiment also demonstrates the general point that voice-like noise ('formant noise') can be interpreted as actual voices.

Summary

Our brains interpret speech by recognising specific frequency peak ratios between different sound frequencies. If there are noticeable peak frequencies, in what is otherwise random noise, they may be interpreted as formants. Provided there is another peak present with roughly the correct frequency ratio, the result could be interpreted as a phoneme. Noise reduction software can accentuate existing peaks in random noise. This increases the chance of frequency peaks appearing as phonemes and also creates time gaps that can be interpreted as spaces between words. By also filtering out the higher frequency sound, a more compact spectrum envelope is produced which will encourage our brains to interpret the noise as speech. This effect works best at the lowest frequencies. By a 'top-down' process, our brains impose a coherent meaning ('words') onto the apparent phonemes, producing relevant 'messages'.

Thus can meaningless noise, provided it is not completely random (like pure white noise), be transformed through software processing into 'formant noise'. This 'formant noise' can be interpreted, by some people sometimes, as words and phrases. Therefore, such audio editing is to be avoided, if possible.

Important Note: You may be filtering sound and applying a spectrum envelope without realising it and without using any editing software. If you set your recorder to a low quality speech mode, the frequency response may be 3000 Hz or less. You should select and use sound recorders with care.

Experiment 2: More sound editing

Listen to sample 1 - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Here is a frequency graph of the sample:

Frequency graph of sound sample

In this graph, frequency is shown along the bottom, increasing rightwards. Time is along the right, increasing into the distance and the vertical scale is sound volume. This is an alternative way of displaying sound frequency to the spectrograms given above. Note how the peaks look isolated and are composed of, more or less, regular triangular shapes.

Now listen to sample 2 - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

How the clips were produced

Spoiler warning! You should do the experiment before reading this section! The two clips in this experiment started out as real, ordinary speech (with large gaps between the words to improve clarity) against a loud background of white noise. This was then processed with noise reduction software (using heavier settings than in experiment 1) to remove the background noise. The result was sample 1. The software certainly succeeded in removing the noise, leaving an uncanny silence between the words. The point to note is that the noise reduction also affected the real voice. Would you recognise who was speaking from these samples?

The second clip was the same noise-reduced recording but with every frequency over 2000 Hz filtered out as well. This had the effect of removing much of the sound of the consonants. Consonants tend to have higher frequencies than vowel sounds. Consonants increase the intelligibility of speech so degrading them in this way can only make the words harder to follow. Listen, for instance, to the first word in sample 2?

Discussion of Experiment 2

Spoiler warning! You should do the experiment before reading this section! The point of this experiment was to explore what happens when you use heavy noise reduction to 'improve' or 'clean up' EVP samples.

The first sample sounds remarkably like sine wave speech. In sine wave speech, the formants are represented by pure sine waves without all the minor harmonics that contribute to the overall sound of natural speech. Given that noise reduction suppresses minor sound variations and accentuates major ones, it isn't too surprising that it ends up sounding like sine wave speech. Effectively all the individual character and tone of the voice has been stripped away leaving just the bare formants. These can still be understood as words because our speech recognition system depends heavily on formants (with interpretation of actual content often influenced by context and expectation).

The second sample has had all the frequencies over 2000 Hz stripped away. This degrades the consonants which normally increase intelligibility. The first word, in particular, has been reduced to a mere burst of noise in sample 2. The phoneme restoration effect (which was operating in experiment 1) can't help this time because that requires white noise in the background to work. Thus, even with repeated listenings, many people will not hear any alternative interpretations to the one we know it to be. In experiment 1, by contrast, with lots of noise in the background, all sorts of interpretations were possible.

Once again, the message is to try to avoid audio editing to 'clean up' your recordings. These methods can actually degrade, or substantially alter, ordinary speech. If the same methods appear to 'improve' EVP clips then you have to seriously consider that you may not be dealing with a real voice at all but just an artifact of the software processing.

Experiment 3: Voice content

Listen to sample 1 - (rather faint recording) set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Now listen to sample 2 - (rather faint recording) set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Voice at increasing distances

The figure above shows the frequency spectrum (a 3-D spectrogram) as the word LET is said out loud (ordinary human speech) at increasing distances from the recorder. Time is the bottom axis, increasing rightwards (0 to 28s). The right axis is frequency, increasing from the back (from 45 to 5000 Hz). The vertical axis is loudness. You can see that the word LET is said six times in all (the parallel lines of peaks coming towards the front like chains of mountains poking out of light blue clouds). When the word LET is first said (far left), there are obvious high peaks - the formants - up to 5000 Hz (good enough for voiceprints!). As the speaker moves further away (going right), the number of formants gradually diminishes. They also shrink in height until they are no higher than the low frequency background noise forming a 'wall' at the 'back' of the graph. The higher frequency formants (above about 3000 Hz) vanish completely. In addition, the lowest frequency formants vanish into the background noise. Speech recorded at a distance will have few formants, mostly in the region below 3000 Hz.

How the clips were produced

Spoiler warning! You should do the experiment before reading this section! This time a real voice was used, albeit in the distance! Sample 2 is the unprocessed version with lots of background noise. Sample 1 has had all frequencies above 2000 Hz removed.

Discussion of Experiment 3

Spoiler warning! You should do the experiment before reading this section! You probably found that sample 1 appeared clearer than sample 2 because the intrusive noise was reduced. However, you may have noticed that with sample 2 that your interpretation of the words did not seem to drift as much, even when you saw alternatives.

By using real words, the possible interpretations tended to be restricted compared to experiment 1. Though you might hear different interpretations in sample 1, they probably all sounded essentially similar.

The words are hard to make out because they are said at a distance. This means that many of the higher frequencies have vanished which degrades the consonants in particular (see figure above). Deliberately filtering frequencies above 2000 Hz makes the recording sound clearer but it actually is more difficult to understand. That's why you may have noticed that the words might have appeared to drift more between possible interpretations in sample 1 than in sample 2. The marginally better consonants in sample 2 restricts the possible interpretations.

The point of this experiment was to see how distant real, ordinary voices could produce 'formant noise' which could be interpreted as paranormal voices. The message content was restricted by using a real voice but nevertheless drifted from the original words. The real words spoken were actually 'ducks high', though you probably can't make that out. The two words were deliberately chosen to be an unlikely sequence. Contrast that with the possible interpretations below which are all likely, if rather meaningless, word combinations.

From a wider perspective, this experiment demonstrates that distant voices, not noticed at the time of recording, can be re-interpreted with different content. This content can be influenced by suggestion so that it appears to be relevant to the situation where the recording was made (eg. a message from a ghost). The phoneme restoration effect probably has a big influence in sounds like this which are based on real voices.

Experiment 4: Whispering

Listen to sample 1 - (rather faint recording) set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Here is a frequency graph (3-D spectrogram) of the sample:

Spectrogram of whispering sound

This figure shows time (bottom, increasing rightwards) against frequency (right axis increasing from back from 45 to 4000 Hz) and sound volume (vertical axis). Note how most of the sound is in the very low frequency range (below 500 Hz) while there are minor peaks arranged like mountain ridges at higher frequencies. Compare this to previous figure.

How the clip was produced

Spoiler warning! You should do the experiment before reading this section! This clip was completely unprocessed. A noise filter was tried but it ended up sounding metallic and less clear than the original! That's probably because the background noise, which is used by the phoneme restoration effect, was removed.

Though it sounds like someone whispering, the sound is actually someone shifting uncomfortably in a wooden seat, just as they might do in a long vigil! The squeak you hear is from the chair. The rest of the noise (the 'whispering') is the sound of clothes rustling.

Discussion of Experiment 4

Spoiler warning! You should do the experiment before reading this section! Whispering is frequently reported in haunting cases. If you replayed a tape from a vigil and heard this sound you might be tempted to say it was ghostly whispering. However, this recording was deliberately manufactured by recording someone shifting about in a chair!

Whispering is a degraded version of normal speech. Voicing (the low frequency sound made by the larynx) is completely absent, vowels lose their strong F1 formant frequency and pitch variation vanishes. It is difficult to make out who a person is when they are whispering as accents and other minor variations disappear. When your brain hears whispering, it has to work extra hard to understand words and frequently makes mistakes. Indeed, it may 'hear' whispering where none exists, as in this example.

An interesting point to note here is that all the possible interpretations of the 'whispering' are similar sounds (compared to experiment 1, for instance). The reason for this is those 'ridges' you can see in the spectrogram above. They show significant high frequency components at intervals in the sound clip. Your brain interprets these as consonant sounds which provide timing for possible syllable layouts. This restricts the possible interpretations to certain syllables and explains the particularly meaningless 'messages'! They are the best the brain can come up within the restrictions provided by the consonant-like sounds.

The point of this experiment was to show how certain common sounds from vigil recordings could appear as whispering. The human brain lowers its standards of speech recognition for whispering, so you should treat all such recordings very carefully. It could just be the sound of someone moving about in their seat! You may be able to deduce this from the wider context eg. were there other more obvious sounds of shuffling and chair creaking before and after the suspected whispering? It is difficult to remember all the small, insignificant things that happen on a vigil so you should treat such marginal samples as suspect and examine them rigorously.

Experiment 5: Paranormal sounds

Listen to sample 1 - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Listen to sample 2 - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Here is a frequency graph (3-D spectrogram) of sample 1:

Paranormal sound graph

This figure shows time (bottom, increasing rightwards) against frequency (right axis increasing from back from 45 to 2500 Hz) and sound volume (vertical axis). Note how the sound forms regular lines of strong peaks. Also, note the lack of any sound above 2000 Hz.

Listen to sample 3 - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Experiment 5 sample 3 spectrogram

This is a spectrogram of sample 3. Notice the lack of any much sound below 800 Hz or above 2100 Hz. Notice also the strong isolated peaks.

Listen to sample 4 - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Listen to sample 5 - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Listen to sample 6 - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Listen to sample 7 - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Listen to sample 8 - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Listen to sample 9 - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Listen to sample 10 - set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Listen to sample 11 set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Also, compare sample 11 with sample llA.

How the clips were produced

Spoiler warning! You should do the experiment before reading this section! Sample2 is a recording of radio interference from a cell phone near the recorder. You cannot hear any actual conversation as the signal is digital and encrypted. Sample 1 is a processed version where, first noise removal has taken place and then, a filter has been used to exclude all frequencies above 2000 Hz (see diagram above).

Sample 4 is someone crinkling or rustling paper in their hands near the recorder. This could certainly happen in a vigil (someone fiddling with their notes, for instance) without anyone noticing it at the time. Sample 3 is a processed version where, first noise removal has taken place and then, a filter has been used to exclude all frequencies above 2100 Hz. There was definitely NO real voice recorded!

Sample 6 is the sound of someone carrying a sound recorder in their hand. The enigmatic noises are mostly transmitted within the recorder case (rather than through the air) and would not be noticed at the time of recording. Nobody sighed or spoke at all during the recording. Sample 5 is a processed version where, first noise removal has taken place and then, a filter has been used to exclude all frequencies above 2000 Hz.

Sample 8 is someone moving crockery and cutlery in another room, a floor down! It has been amplified but otherwise is unprocessed. Sample 7 is a filtered version which has excluded all frequencies above 2000 Hz.

Sample 9 is sample 3 played in reverse!

Sample 10 was deliberately constructed in audio editing software using bits of white noise. They were set at intervals that resemble the rhythm of real, normal speech from recordings. There was no attempt to replicate formant frequencies though they may be present by pure chance. To enhance the 'voice-like' quality, the whole recording was filtered to exclude all frequencies above 3000 Hz.

Sample 11A was produced with a sound recorder using an internal microphone. The noise comes from handling the machine during recording. Sample 11 was an extract from sample llA, featuring just a brief section and with all frequencies above 2000 Hz filtered out. It can be difficult to pick out sample 11 from llA. It is right at the end, if that helps.

Discussion of Experiment 5

Spoiler warning! You should do the experiment before reading this section! The point of this experiment was to see how many natural sounds, that could be heard on a recording made on a ghost vigil, could be identified afterwards. Clearly, with no visual record it can be difficult to decide, after the event, what each sound might be and whether it is really paranormal or not.

Samples 1 and 2 are mobile phone radio interference. Unprocessed, the sound is distinctive and obvious. After processing, some people may have heard a curious 'singing' sound in the background. There is no such sound obvious in the original recording (sample 2) so it must be an artifact of the noise processing and filtering. The noise removal process, in particular, tends to produce pure tones which could account for the musical quality of the background.

The 'voice' (or 'voices') in sample 3 is so distinct that most people are likely to hear them, particularly if the sample is repeated and played loudly. Many people may even agree with one or other (or even both!) of the interpretations given. The strong impression of separate words is probably a result of the big variations in intensity caused by the rustling. This was accentuated by the noise reduction process. The quite plausible 'voice' is probably helped by the tight spectrum envelope between 1000 and 2100 Hz. The paper rustling did not produce much really low frequency sound. These higher frequencies give the impression of consonants which, in normal speech, aid intelligibility. The fact remains, however, that the sound consisted solely of rustling paper (derived from sample 4)! The lesson is, you should keep your sound recorder well away from anything that might be knocked, pushed, rustled or brushed by investigators!

Samples 5 and 6 demonstrate why you should not hold your recorder in your hand while recording. The noises produced are loud and ambiguous and drown out other, potentially more interesting, sounds. They could be interpreted as various things after the event, including physical ghostly activity. Processing the noise changes it without making it any clearer what the real source of the sound is.

In sample 8 the background noise (a white noise hiss) is prominent, indicating that the main sound (the clashing crockery and cutlery) is not much louder. It is amazing that a sound like that can be picked up so well from such a distance. Sound recorders can be quite sensitive, more so than humans. It is unlikely that anybody on a vigil would have noticed this faint sound at the time of the recording. The curious rhythmic quality is just a coincidence. It is part of a longer sequence of such noises. By isolating one part of this sequence of sounds and listening to it repeatedly, it gives the false impression of a regular metallic mechanism. That is why you should always include several seconds before and after an apparently anomalous sound, in a clip, to give the listener the context of the recording. In sample 7 the higher frequencies, which are prominent in this sort of noise, have been filtered out. It gives the sound a more voice-like quality though it is still difficult to make out what might be said.

Sample 9 once again demonstrates how playing voice-like noise backwards can still produce apparently perfectly intelligible messages. Note that there is no 'second voice' in this recording, despite the fact that it is sample 3 played backwards (which contained an apparently quite distinct 'second voice'). Noise is always just noise whichever way round you play it, unlike real voices.

Sample 10 was all about seeing how important rhythms are, compared to formant frequencies. The result was surprisingly effective, though before it was filtered it sounded much more like the bursts of white noise it really is. It is clear that the human brain uses both formant frequencies AND timing to interpret a sound as speech. The very hurried delivery of apparent words in this example may be because, in real speech the sounds build gradually in intensity and then decline while, in this sample, they appeared abruptly at full volume. No doubt with a bit of fiddling around, a more convincing sound could be produced.

Sample 11A shows the sort of noise generated by handling a microphone, or a recorder containing one, while recording. Even in the unedited 11A, the sound can be remarkably voice-like in quality (does it say 'winter' at the end of the burst of noise?). By isolating a short section towards the end (sample 11), which happens to have rhythms similar to talking, and filtering out the distracting higher frequencies, it sounds even more like words. Even without filtering and noise reduction, simply isolating a short stretch of sound from its context can make it sound more like a voice. Context is important in telling our brains whether they are to expect noise or voices and so can affect how extracts are interpreted. Note, also, how the apparent words change with the length of the recording.

Experiment 6: Low frequency

Listen to sample 1 - (quite a faint recording) set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Listen to sample 2 - (a very faint recording) set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

Listen to sample 3 - (a very faint recording) set your audio software to repeat, or continuous play, to try to catch the words - once you've decided what you think it's saying, follow this link.

How the clips were produced

Spoiler warning! You should do the experiment before reading this section! Sample 1 contains a human voice derived from the same source as Experiment 2 Sample 1 - it is saying 'this is a message'. All the frequencies above 300 Hz have been filtered out. Sample 2 is derived from Experiment 5 Sample 4 (rustling paper). Again, all the frequencies above 300 Hz have been filtered out. Sample 3 is derived from Experiment 1 Sample 2 with the same treatment as the previous two samples.

Discussion of experiment 6

Spoiler warning! You should do the experiment before reading this section! It has been suggested that voices heard below 300 Hz may be paranormal because this is the lower limit of the natural human voice. As sample 1 demonstrates, some of a normal voice can still be heard below 300 Hz. It can even be intelligible!

In sample 2, virtually the entire sound of the rustling paper was removed when everything above 300 Hz was filtered out. In sample 3 a previous sample that sounded like a voice (click here for the original) has been treated in the same way. In this case, however, you can still hear the apparent voice, albeit rather badly. What this shows is that formant noise, as well as real natural voices, can still be heard below 300 Hz.

So, the fact that part of a recording of an apparent voice may dip below 300 Hz does not automatically make it paranormal.

Postscript to experiments

Recent research in the field of synaesthesia may suggest a reason why we can sometimes hear 'words' in otherwise meaningless noise. Synaesthesia is a condition, shared to some extent by around 1 in 20 of the population, where people unconsciously form seemingly bizarre connections between various senses. They may see music, smell shapes or hear colours. This is fascinating enough in itself and may explain some reports of apparent paranormal phenomena. However, the really interesting new discovery is that the area of the brain that allows the senses to link in this way is the one that processes language - the bit where concepts are linked to words. It seems all sensory input may be processed through the language areas of the brain. What is more, one kind of sensory input can affect interpretation of another. Moreover, the connection synaesthetes make between sensed objects are not random. There are conceptual connections linked through language.

Though most people are not synaesthetes, it is quite possible that anyone, faced with ambiguous sensory input (like 'formant noise'), may have their experiences 're- interpreted' for them by the language area of their brain. This could explain optical and aural illusions where, what we see or hear, is affected as much by what we think as by the sensory stimulus. It would also explain the role of suggestion, where thoughts, or simultaneous sensory inputs, affect what we experience (eg. the interpretations in experiment 1 changing what was 'heard').

 

White noise

True white noise. It is spread over all sound frequencies.

See also

Additional EVP gallery here.

Notes (do the experiments before looking here!)

If the link hasn't led you to precisely where you expected, use the name of each paragraph below to navigate. Thus, if you want the commentary on the second sound in experiment 1, go to 'Experiment 1 Sample 2'. You may disagree with the comments made here. This only goes to show that listening to clips like this is subjective. There are no right or wrong answers as it is an experiment!

Experiment 1 Sample 1: It appears to say 'We're the lost people'! Or is it ''On that steeple' or ''Oh yes we do' or ''You're a student' or 'I have a secret' or 'The long lost treacle' or 'Pleased to see me' or even 'Tone deaf seagull'! Or did you hear something completely different or no voice at all, just a noise? Try listening again (with repeat or continuous play on) while reading each of these interpretations, once at a time, in turn. Also, try listening with and without significant background noise. If you keep listening to the recording repeatedly in a loop, you will probably hear the words change without prompting. This is the verbal transformation effect. Return to experiment.

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

 

Experiment 1 Sample 2: It is essentially the same as the previous sample ('We're the lost people', 'oh yes we do', etc) but more difficult to make out. It sounds more hissy. You may hear the same words, or something different or maybe you can't make out words at all. Return to experiment.

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

 

Experiment 1 Sample 3: This will probably sound like white noise to most people who will be unable to make out any words. You may still hear what you heard before but only very indistinctly. You might possibly catch the word 'yep' instead! Try this sample first on someone else, who hasn't heard the previous samples, and see if they can make anything out. Return to experiment.

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

Experiment 1 Sample 3c: It appears to say 'A beast to work'! Or is it 'At least it works'' or 'A feast similar' or 'The peace seminar' or 'The good southern air' or even ' A goose over there' perhaps. Or did you hear something completely different or no voice at all, just a noise? Try listening again (with repeat or continuous play on) while reading each of these interpretations, once at a time, in turn. Notice the similarity in tone to sample 1. Return to experiment.

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

Experiment 1 Sample 4: This sample (4) contains sample 3 though you may have difficulty pinpointing exactly where it is. By hearing a longer sample you find any 'voice' you may have previously heard in sample 3 sounds less convincing. On the other hand, you may hear a new 'voice' in sample 4 which is spread out over much of the longer clip (do you hear the words 'in again, out again and out again' for instance). If you listen to sample 3 again straight after, you may find the 'words' of this clip have changed to a much shorter 'message' (a small part of the 'words' you may have heard in sample 4, in fact - just the final 'out again' maybe) ! Or you may simply think that both are just noise. If you listen to sample 1 again, it probably still sounds the same as when you first heard it ie. it was not influenced by hearing sample 4 or 3. Despite the fact that sample 1 is derived from sample 3 they sound completely different! If you had made these recordings yourself, which would you say was correct - the processed or unprocessed sound? Sample 3 is towards the end of sample 4, in case you need a hint. Return to experiment.

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

 

Experiment 2 Sample 1: This weird sounding voice is saying 'this is a message'. It's not too difficult to understand with a couple of repeats. It sounds very like sine wave speech. It is so clear that even repeated listening is unlikely to change the interpretation. Return to experiment.

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

 

Experiment 2 Sample 2 : To some people this may sound clearer than the previous sample, to others not so clear. As you will have no doubt worked out, it is the same message. It sounds less hissy and 'metallic'. By listening to this clip repeatedly, you may hear other interpretations of the message. Can you hear 'Disc is a missive', for instance? Return to experiment.

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

 

Experiment 3 Sample 1 : It appears to say 'That's fine' or ' Dutch pine' or 'Lunch pie' or maybe even 'Checkpoint'. Or maybe you heard something else or just noise. Return to experiment.

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

 

Experiment 3 Sample 2 : This is the same recording as sample 1. Does it sound clearer? There is certainly a lot of annoying hissing noise which doesn't seem to help. Can you make out the words any better? On repeated listenings, do the words change? Try reading all the suggestions above for experiment 3 sample 1. Can you hear any of them? Return to experiment.

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

 

Experiment 4 Sample 1 : The whispering voice is very hard to make out (as are all whispers). It sounds a bit like a squeak followed by 'half mast hunger' or ' half price plunder' or 'farm slice under' or 'pass asunder' or 'dark as thunder' or even 'purr so fine'. Try listening again (with repeat or continuous play on) while reading each of these interpretations, once at a time, in turn. No doubt there are other variations on this theme that you can hear as well. Return to experiment.

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

 

Experiment 5 Sample 1 : Ignoring the urgent 'hammering' sound in the foreground, do you hear someone calling or singing in the background? It is too difficult to make out what is being sung but it definitely sounds like a voice. Return to experiment.

 

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

Experiment 5 Sample 2 : This is an unprocessed version of the previous sample. Many of you will now recognise what it is, if you hadn't already. Do you still hear any singing? Return to experiment.

 

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

Experiment 5 Sample 3 : Many people will hear a distinct voice here above the odd 'sing-song' background noises. Maybe it is saying 'Wolf is arrogant fellow'. It sounds as though a second voice is chiming in on the final 'fellow'! Or maybe you hear 'Youth in Harrogate, hello' (again with a second voice chiming in on 'hello') or maybe 'Who is daddy's girl?'. Or you may hear other variations or maybe just noise. You may well find it clearer and less ambiguous than previous samples. Note the strange rhythm and odd changes in intonation. Return to experiment.

 

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

Experiment 5 Sample 4 : This is an unprocessed version of the previous sample. You will probably not hear a voice this time (though there may be a hint of a loud undecipherable whisper). The 'second voice' at the end is still voice-like though it is difficult to make out what it is saying - possibly 'heaven'. Have you guessed what the sound really is? Return to experiment.

 

 

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

Experiment 5 Sample 5 : This sounds like someone dragging heavy furniture about and then the word 'hello' is whispered. This sounds like an interesting recording for a haunted house! Return to experiment.

 

 

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

Experiment 5 Sample 6 : This sounds like something being scrapped over a rough surface followed by a loud sigh! The sigh could also be interpreted as an enigmatic 'Tuh!' or even 'Tello' (whatever that means). Sighs have often been reported on vigils. Return to experiment.

 

 

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

Experiment 5 Sample 7 : This sounds like something mechanical - some kind of ghostly old-fashioned machinery, perhaps, like a large clock mechanism. It is difficult to make any words out unless they are said in a curious slow sing-song way. Maybe 'that spiral please' perhaps. Return to experiment.

 

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

Experiment 5 Sample 8 : This is clearly the same sound but with more hissy background noise evident. It could still be some sort of machinery but it has more of a metallic quality now. Difficult to see any kind of voice in this sound. Return to experiment.

 

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

Experiment 5 Sample 9 : Many people will hear a distinct voice here above the odd 'sing-song' background noises. Maybe it is saying 'Both the car gears were frozen'' or perhaps 'Vote for July cheese whenever'. Or you may hear other words or maybe just noise. Note the strange rhythm and odd changes in intonation. Return to experiment.

 

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

Experiment 5 Sample 10 : It sounds like a very hoarse voice saying 'it matters' or 'I'm nervous'. The words are said very quickly. Some people may just hear noise. Return to experiment.

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

Experiment 5 Sample 11 : If listened to repeatedly you may here 'hello' or 'bizarre' or even 'reach out' or 'breathe out' or in a strange hoarse tone. Some people may just hear noise. Return to experiment.

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

Experiment 6 Sample 1 : If you keep listening to this recording you should recognise that words are being said. They sound heavily muffled and distant. They are saying 'this is a message', which you will probably pick up once you know. Return to experiment.

 

 

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

Experiment 6 Sample 2 : Difficult to hear anything here. It sounds a bit like the wind with a dull mechanical noise towards the end. Return to experiment.

 

 

^^^Read the text at the TOP of the page ONLY^^^

Spoiler warning - do not read beyond here!

 

 

 

 

Experiment 6 Sample 3 : Difficult to hear anything here. However you may be able to make out a muffled version of 'We're the lost people'! Or is it ''On that steeple' or ''Oh yes we do' or ''You're a student' or 'I have a secret' or 'The long lost treacle' or 'Pleased to see me' or even 'Tone deaf seagull'! Return to experiment.

Pure electronic white noise spectrogram

This is 'real' white noise. It is evenly spread over all frequencies.

Radio white noise

This is 'radio' white noise ie. the noise you find when 'tuned' between radio stations (in the spectrogram, frequency increases downwards, time rightwards). Unlike real white noise, it is not evenly spread over all sound frequencies. Note the peaks around 1500 Hz and 3000 Hz.

Note: All the sound clips on this page were deliberately recorded and processed for the purposes of demonstration. They were not recorded at haunted, or any other special, locations.

© Maurice Townsend 2007 , 2011, 2012