Episode: 4658
Title: Audio Revisited
Source: https://hub.hackerpublicradio.org/ccdn.php?filename=/eps/hpr4658/hpr4658.mp3
Transcribed: 2026-07-31 16:16:07 (official HPR transcript)

---

This is Hacker Public Radio Episode 4658, for 2026-06-10
Today's show is entitled, "Audio Revisited"
The host is Whiskeyjack and the duration is 00:50:23
The flag is Clean, and the license is CC-BY-SA
The summary is "This is a follow up to the series on Simple Podcasting showing the results of filtering experiments"
This is a follow-up to my four-part series on simple podcasting.
In this episode, I will discuss a number of experiments with audio filtering.
These experiments were inspired by comments by listeners and by other discussions about
audio on HPR.
I am not an audio expert, so I am doing this partly in order to learn something, but mainly
in order to have a bit of fun.
I hope that you find this entertaining as well.
In a comment on the first episode, a listener mentioned something called solo cast
and said that the method bore resemblance to the method that I was using.
Here is this comment.
Comment by Rito.
It reminds me about solo cast, high whiskey jack, I really liked your podcast and the topic.
I cannot remember about your last, but the sound quality of this one was good in my mobile
speakers.
The concept reminded me about the program from Norris T, another host on HPR, while similar
does it have some differences in HPR 3496, and then it gives the URL to the episode.
As I am not on the future feed, I look forward to your next episode.
Here is Rito.
End of comment.
I did not recall having heard the episode on solo cast, but this sounded very interesting.
Solo cast was an HPR episode 3496 and was released by Norris T on the 20th of December 2021.
I listened to that episode and it does indeed use the same basic concept of recording
short segments of audio and combining them later instead of creating one big recording and
editing it with an audio editor.
The main difference is that the workflow that I described involves a lot of manual steps,
while solo cast is a short Python program that automates the entire process of presenting
your script, recording the segments, combining the segments, and filtering and normalizing
the result.
I won't try to describe solo cast in detail instead I would recommend just listening
to the HPR episode 3496 to get Norris T's explanation directly.
Well, I wanted to make sure that a credited Norris T with having come up with this concept
four years before I did, this won't be the focus of this episode.
Instead, I will talk about audio filtering and various experiments that are ran on different
methods.
While looking at the source code for solo cast, I noticed that it used a filtering method
that resembled one used by JiveTalk, a podcast production program that caught the attention
of one of the HPR community news presenters.
This method involves taking a sample of quiet audio where there is no speaking taking place
and then using this as input to a noise reduction filter which is applied to the voice recording.
The filters subtracts the quiet sample from the voice audio which should theoretically
remove the ambient noise.
I decided to apply this method to a number of different audio test recordings which were recorded
under different circumstances using different hardware.
In this way, I could see if the method worked equally well under all circumstances, or
if there were some sorts of noise which it was suited to and some sorts that were not.
While I was at it, I also picked several other filter methods to see how they worked as
well.
Potentially, some methods may be better under some conditions while other methods were
better suited to others.
I won't present all of my experiments as that would be a bit dull to listen to.
Instead, I will describe each method and then present audio samples which illustrate
my conclusions.
There are two pieces of audio software involved, both of which were also used in my series
on simple podcasting.
The first is Sox, spelled SOx and which is short for sound exchange.
Sox is a command line program for audio manipulation.
Sox is free software, released under the GPL V2 or later.
The other is FFMPEG, which is also a command line program.
FFMPEG is also free software, released under the LGPL V2.1 later and GPL V2 or later.
Sox actually uses FFMPEG for certain operations.
For recording hardware, I use the following.
Maxwell Headset, the first is a cheap Maxwell Headset that has an electrical noise problem.
Unfortunately, I don't have a model number for this headset.
I described this hardware, the noise problems that I had with it and how I created filters
to deal with the noise in my series on simple podcasting.
Briefly though, this is a headset that has a built-in microphone on a boom, which allows
the microphone to be positioned close to the mouth.
It connects with a USB cable.
Born, earpiece and inline microphone.
This is a set of earplugs that go in your ears and connect it by wires and a very small
microphone built into a small bulge in the cable.
It connects using a 3.5mm jack.
The model number seems to be BUD250-BL.
X-Tryk headset.
This is a gaming headset similar to the Maxwell headset described above.
The model number is GH-510.
It uses a USB connection.
Yann May condenser microphone.
This is a microphone that comes with a small tripod stand.
The model number is SF-910.
It uses a 3.5mm audio jack.
This is not a review of the hardware.
Rather, I was trying to create audio problems so that I could test ways to fix them.
Therefore do not take the above list as a recommendation of what to buy.
However, you can see that I am not using any expensive audio hardware.
If you want to make an HPR podcast, you do not need professional level hardware.
Audio samples.
The audio samples are as follows.
Quiet.
This was recorded in a quiet environment at my desk.
This is my normal podcasting environment and represents optimal conditions.
The main reason for this method is to see how the various filter methods perform when
dealing with electrical noise from the Maxwell headset.
Small fan.
This is a small USB powered table fan approximately 10 cm in diameter.
It was located roughly 40 cm or less to the left of the microphone, although this varied
depending on the microphone.
Traffic.
This was along a busy street with traffic noise in the background.
Filter methods.
Socks.
Noise RED filter with audio profile.
This method uses the Socks.
Noise RED or N-O-I-S-E-R-E-D filter.
Here is a brief quote from the Socks documentation on this filter.
Quote.
Reduce noise in the audio signal by profiling and filtering.
This effect is moderately effective at removing consistent background noise, such as his or
hum.
To use it, first run Socks with the Noise Prof that's N-O-I-S-E-P-R-O-F effect on a section
of audio that ideally would contain silence, but in fact contains noise.
Such sections are typically found at the beginning or end of a recording, end of quote.
For these tests, I recorded a separate noise profile to go with each test.
Basic manual filter.
This is a basic, high and low pass filter pair based on the work I had done in my previous
series on simple podcasting.
However, based on the tests that I have done for this episode, I decided to get a bit
more aggressive in terms of filtering.
I use a high pass filter of 120 Hz and a low pass filter of 8 kHz.
Each filter is then applied twice to increase the its effect.
I also added band reject filters to deal specifically with 50 and 60 Hz line noise.
Complex manual filter.
This uses the manually constructed filter described in my series on simple podcasting.
This uses the basic manual filter plus a series of custom band reject filters to fix specific
noise problems in the Maxwell headset.
F-F-M-P-G-A-F-F-T-D-N filter.
The documentation describes this as D-noise audio samples with F-F-T.
F-M-P-G-A-R-N-D-N filter.
The documentation describes this as reduced noise from speech using recurrent neural networks.
F-F-M-P-G-A-T-E filter.
I will pronounce this as A-G-T for convenience.
The documentation describes this as A-G-T is mainly used to reduce lower parts of a signal.
This kind of signal processing reduces disturbing noise between useful signals.
Method.
The experimental method used was to take each noise sample and apply different filter methods
to it.
For there are parameters which can be adjusted, a script was used to generate a series
of different sample files with different parameter values.
Not all possible parameters were experimented with as the goal is to see which method
produces what sorts of results under different circumstances not to get the best possible
result for the samples that I happen to have.
The method in each case was as follows.
Step 1.
Convert the audio file to FLAC if it is not already in that format.
Step 2.
Apply basic HI and low-pass filter described previously to each sample.
The reason for this basic filtering is that it eliminates at least some undesired noise
in a fairly full-proof manner, leaving less for the more advanced filter to deal with.
This should allow for a better test of the filtering under realistic conditions.
Step 3.
Apply the noise reduction filter being tested.
Step 4.
Normalize the filtered sample to 17 LUFS according to the EBU R128 standard.
The EBU standard is described in my series on simple podcasting.
Normalizing adjust the audio signal to the desired loudness cell level.
This allows for more consistent sound levels and allows us to hear the results under realistic
conditions.
I normalize the audio individually for each sample as different recording hardware requires
different amounts of loudness adjustment.
This is different from the typical podcast process where normalizing takes place at the
very last step in the process, but it was necessary in this case.
Step 5.
Concatenate selected sample audio files to one another to allow for better review and comparing.
Results.
The results are grouped according to the type of noise which is being mitigated.
This allows for easier comparison of the effectiveness of each technique under different
circumstances.
I've only picked a few examples of interest out of the numerous experiments that I conducted.
Quite recording environment with Maxwell headset.
This compares how well various filtering methods work on the noise induced by the electronics
in the Maxwell headset.
This electronic noise consisted of noise spikes every one killer hurts.
This should be representative of electronic noise caused by problems in recording hardware.
Manual filter.
The manual filter, applied in narrow, band-reject filter, every one killer hurts from one
killer hurts to 12 killer hurts.
Just completely remove the otherwise audible wine caused by the noise.
FFMPEG, AFFK, DN.
This method allows for setting a noise floor and specifying how much the noise floor should
be reduced by.
The method is very sensitive to getting the noise floor correct for that recording.
Set the floor to low and nothing happens.
Set it to high in some distortion results.
However, it seems to be moderately effective, but it would seem to require checking it
and possibly adjusting it each time it is used.
FFMPEG, AFFMPEG, A8.
That's AG, ATE.
This method allows setting a noise floor and suppressing all sound which falls below that
level.
This method is very sensitive to getting the noise floor correct for that recording.
If set to low or quiet, it is ineffective.
If set to high or loud, it distorts words which come after a pause, which would typically
be between sentences.
When set correctly, it completely removes noise in the silences between sentences.
However, the noise is still audible during speech.
This is because the noise in this case is a higher frequency than the normal speech and
so stands out more.
It may not be a significant problem for the noise which is closer to the main vocal frequency
band.
Overall, this method is not suitable for this particular problem.
FFMPEG, ARNN, DN.
This method uses a standard model.
A variety of different noise reduction models are available.
I tested it with only one, STDN, dot R, and NN.
It does not seem to induce much distortion in the voice signal, even with a high amount
of mixed parameter.
However, it is only slightly effective at removing the line from the signal, even with
a high amount of mixed parameter.
Overall, this method does not appear to be useful for this sort of noise problem.
Socks, noise, ARND filter.
This was effective in removing noise between words, but noise can be heard while words
are being spoken.
It was better than a gate, however.
Overall conclusion for the Maxwell headset noise.
When dealing with narrow noise bands that occur at known frequencies, the manual filter
is leagues ahead of any of the other tested alternatives.
Sample Audio.
Here is a sample audio recording showing the best overall results.
The sample is repeated, first with only basic low and high-pass filtering, and then
with the manually constructed filter.
In the first sample, you should hear a high-pitched background wine, and the second sample,
the high-pitched wine is completely removed.
This is a test, 1, 2, 3, 4, 5.
This is a test, 1, 2, 3, 4, 5.
Traffic noise.
This was recorded using the born in line microphone connected to a mobile phone while walking
along beside a busy street.
This was in dry, cool, spring weather, and the road was paved with asphalt.
This should be reasonably representative of podcasting while walking outdoors in an noisy environment.
Basic manual filter.
This used the basic manual filter with high and low-pass filters.
This did nothing very useful in this case, as the signal was already filtered within
those limits by the recording hardware anyway.
The low sample rate of 8 kHz in the phone limited the upper frequency to 4 kHz.
The sample rate has to be twice the highest frequency that you want to detect.
Overall, this is not suitable for this sort of problem.
FFMPEG AFFT-DN.
With a high noise floor, background noise is reduced, but not eliminated.
There was not much distortion in the voice.
This is only slightly useful for this sort of problem.
FFMPEG A8.
With the high threshold, background noise is reduced, but not eliminated.
There was some distortion in the voice.
The background noise could be heard also heard when speaking, but because the frequency
of the background signal was similar to the louder voice signal, it was not as noticeable
as it would have been if the two were very different.
This is moderately useful for this sort of problem.
It may be more useful in situation where the background noise was not quite as loud.
FFMPEG A8-R-N-N-DN.
With high amounts of noise reduction, much of the background noise is suppressed,
but there is not a lot of distortion in the voice.
The background traffic noise is still present, but is significantly less.
This offers only a moderate improvement.
Socks noise are ED filter.
With small amounts of noise reduction, voice is clear, but traffic noise is present
as a very significant continuous warbling sound at the background.
This is no improvement on the original, and in fact, could be seen as making it worse.
With moderate amounts of noise reduction, traffic noise is mostly gone, but there
are still various squeaks present.
Voice is noticeably distorted.
With large amounts of noise reduction, traffic noise is gone, but voice is highly distorted.
This is moderately useful for this sort of problem, but requires careful adjustment.
FFMPEG A-R-N-N-DN followed by FFMPEG A8.
This combined two different filters.
First, it used AR-N-N-DN to suppress the background noise to a lower level without much
voice distortion.
Then it applied the A8 filter to suppress the noise levels between words still further.
This used the same amount of mix and threshold as was found to be most effective when
each of these filters was used on its own.
The background noise is almost completely gone, while distortion of the voice signal is
low.
Overall, conclusion for traffic noise.
The AR-N-N-DN filter combined with the A8 filter was the most successful at suppressing
background noise while limiting the amount of voice signal distortion.
Sample audio.
Here is an audio sample for what I felt to be the best overall results.
The AR-N-N-DN filter combined with the A8 filter.
First is the original audio with basic filtering.
This is followed by the same audio after being passed through the AR-N-N-DN and A8 filters.
This is a test.
This is a test.
One, two, three, four, five.
Another sample.
Here is a second audio sample showing the SOX noise AR-N-D profile-based filter.
I have included this to show how a profile-based filter can make things worse if you are
not careful how you use it.
This repeats the test audio for times.
The first is with basic filtering only.
The second uses low amounts of noise reduction.
The third uses moderate amounts of noise reduction.
The fourth uses high amounts of noise reduction.
The test is the test, one, two, three, four, five.
This is a test, this is a test, one, two, three, four, five.
This is a test, this is a test, one, two, three, four, five.
Spall fan noise with the anime microphone.
This was recorded using the anime condenser microphone.
The small fan was set up behind into the left of the microphone.
This is intended to represent situations where someone may have a fan or air conditioner running
in the background due to hot weather or has allowed computer fan.
A condenser microphone was used for this test as they are more prone to picking up unwanted
noise.
However, for practical recording purposes, this sort of microphone is unsuitable for this type
of environment.
This is a basic manual filter.
This uses the basic manual filter with high and low pass filters.
This did nothing useful as the fan noise was in the same frequency range as the voice signal.
This may be of more help in cases where the noise is below the 120 Hertz cut off used
in the low pass filter.
With high amounts of noise reduction, much of the background noise is suppressed, but
there is some distortion in the voice.
The background fan noise is still present, but is significantly less.
Overall, this is moderately effective.
FFMPEG A gate.
This was effective in removing noise between words, but noise can be heard while words are
being spoken.
However, this was a small voice sample and it is possible that more problems could occur.
With less fan noise than was in this sample, this technique may work much better.
FFMPEG A are NNDN.
With high amounts of noise reduction, much of the background noise is suppressed, but there
is not a lot of distortion in the voice.
The background fan noise is still present, but is significantly less.
Overall, this was fairly effective.
Socks noise are ED filter.
With small amounts of noise reduction, voice is clear, but fan noise is present as a slight
orbling sound in the background.
With moderate amounts of noise reduction, fan noise is gone, but voice is somewhat distorted.
With large amounts of noise reduction, fan noise is gone, but voice is very distorted.
In general, this method is fairly successful at dealing with this sort of problem.
However, there is a trade-off between background noise and voice quality.
Getting that trade-off correct takes experiment and judgment for each specific situation.
FFMPEG A are NNDN followed by FFMPEG A gate.
Discombined two different filters.
First it used AR and NDN to suppress the background noise to a lower level without much
voice distortion.
Then it applied the A gate filter to suppress the noise levels between words still further.
This got rid of virtually all of the background noise between words.
If you listen carefully however, there is a slight buzzing sound in the voice signal.
For all conclusion, for fan noise, with the NMA microphone.
Of the methods used, the AR and NDN followed by A gate filter seemed to offer the most improvement
for the least effort and the least voice distortion.
The AR and NDN filter on its own seemed to be the next most preferable to me, despite leaving
some fan noise in the background.
Audio sample
Here is an audio sample for what I felt to be the best overall results.
The AR and NDN filter combined with the A gate filter.
First is the original audio with basic filtering.
This is followed with the same audio after being passed through the AR and NDN at A gate filters.
This is a test 1, 2, 3, 4, 5.
This is a test 1, 2, 3, 4, 5.
Small fan noise recorded with headset.
The following is an observation, rather than a filtering technique.
When a recording was made using the Maxwell headset and listened to on the headset later,
or with speakers, the fan was virtually inaudible.
When the same recording was listened to with the extra headset, it was barely audible with
careful listening and only detectable as a fan because I knew it was there.
In situations where there is ambient noise, the best noise reduction technique is probably
to move the microphone as close to your mouth as possible, although not directly in front of it,
and reduce the gain if there is a gain adjustment in the microphone.
This will work far better than trying to remove the noise later.
If you are recording an HDR episode at a desk, then an inexpensive headset with boom mic
may do the job just fine with minimal effort and expense.
Conclusions I have tested three noise scenarios, electronic noise in the audio hardware
at specific frequencies, recording outdoors with an inline microphone in a noisy traffic
environment, a noisy fan creating background noise in an office, by conclusion on these
are as follows. Electronic noise in the audio hardware at specific frequencies.
If you can use audacity or some other means to find the frequencies which are causing the noise,
the best solution, assuming you don't just replace the hardware, is to manually construct
filters to remove those specific frequencies. This is a safest solution in terms of only doing
what you tell it to and not producing unexpected surprises sometime down the road when something
changed in the environment. If you are looking for a fairly automatic filtering method,
these socks noise RED profile-based filter seems to work fairly well.
There is the equivalent filter in FFMPEG, but I did not include that in my experiments,
as it is harder to use in a script because it does not use a separate noise profile file.
Recording outdoors with an inline microphone in a noisy traffic environment.
In this situation, the FFMPEG AR and NDN combined with A8 filters seem to be the most successful.
The socks noise RED filter may work, but at the cost of more distortion in the voice
that is seen in the other methods. An inherent problem with any profile-based noise reduction method
is that if the background noise is not constant, which it seldom is in that sort of environment,
the profile may not represent the background noise which is present later on in the recording.
This risks adding more distortion in the voice as the profile and later environments diverge.
However, for this application, a different microphone that provided a better recording
would appear to be advisable. A solution which brought the microphone much closer to the
and so resulted in a better ratio of voice signal compared to background noise
would appear to be necessary after which the question of what sort of noise reduction to use
would then need to be re-evaluated. A noisy fan creating background noise in an office.
The socks noise RED filter and the FFMPEG AR and NDN, AFFT DN, and A8 methods all work to some degree.
However, they all need correct selection of parameters to achieve the proper results.
When I compared all four methods side by side, I found the AR and NDN combined with the A8
filter to be preferable in terms of the trade-off between background noise and distortion of the voice signal.
The AR and NDN filter on its own seemed the next post-prefable to me despite leaving some
fan noise in the background. However, that is a subjective judgment of a specific noise sample
when recorded using a specific microphone. Keep in mind though that many listeners will not
be listening in an ideal environment. They may be doing things where background noise is present
rather than in a very quiet room and so may find a small amount of background noise in the
recording to be less of a problem than distortion in the voice signal which may make some word
harder to understand. When I conducted the same experiment recorded with the X-TryCAD set,
I found that AR and NDN seemed to offer no noticeable improvement. This may be because the amount
of audible fan noise was far less with the X-TryCAD set to begin with. In other words,
there is no single best solution here and you may have to be prepared to try different
options to see which one works best in your situation. The important thing is to
avoid making things worse by applying filtering that is not appropriate for that situation.
The best method may be to use a recording method that doesn't pick up the fan noise to begin with.
This could include just using a gaming headset with boom mic.
I have one final observation on this point regarding headsets. The Maxwell headset has a foam
cover over the microphone, while the X-TryCAD set does not. There was some slight audible wind
buffeting noise picked up by the X-TryCAD set. There was not observed with the Maxwell.
This seemed to offer particular problems with these socks, noise, AR, ED, profile,
base filter, as this noise was irregular and after filtering would show up as a wobbling sound.
If you use a headset and plan to use it in conjunction with a fan, it may be advisable to apply
some sort of wind cover over it. Combining complex filters. In several cases, I found that combining
several complex filters offered better results than using any single one on its own.
The basic strategy though is to first use a method which is good at reducing undesirable noise
without introducing excessive voice distortion. Then apply a different filter which is good at
reducing small levels of background noise to an even lower level while affecting the voice signal
as little as possible. This uses the relative strengths of different filter types to compensate
for the weaknesses of the other. Different combinations of filters were most effective for different
types of problems. I did not try all possible combinations, however. Perhaps a further
exploration of this would be worth doing in a later podcast. K Study, noise in another HPR episode
audio. In the comments to my second episode on simple podcasting, which is HPR 4618, where I discussed
basic filtering a couple of listeners brought up an interesting point. Antoine mentions
deep clicking in a post, Vance replied. Antoine, thanks for mentioning the click removal capability
in Audacity. While I already knew about its noise removal filter, I wasn't aware that it also had
click removal. It might have helped me for HPR 4637, where some sort of electromagnetic signal was
picked up by my microphone and recorder, a Zoom H2. The tapping sound was not present in the room
where I recorded. While click removal does seem to distort speech when applied to it,
though to my ears it doesn't sound as weird as when noise removal is done with speech. I could
have applied the filter only to the pauses, where the tapping sound is most noticeable. I will
consider doing this in the event that I'm not able to eliminate the source of interference in
the future, which would be the best way to go. And of quote, I found this interesting as it sounded
like another audio problem that could be experimented with. I found a sample of the episode which
had the clicks and cut a copy of that segment out to experiment with. These sounds are like a series
of clicks or ticks, which would be another way to describe them in the quiet part of the audio
between sentences or phrases. Next, I use audacity to study the sound spectrum. I found a massive
60 hertz noise spike. However, my speakers won't reproduce sound that low, and filtering this out
didn't reduce the clicks. The clicks turn out to be bursts of noise across the 100 to
800 hertz band, which is right where the main vocal band also is. This makes it difficult to filter
based on frequency. The most promising approach would seem to be to filter based on sound level.
I tried all of the individual audio filter techniques mentioned in the other experiments above.
None produced satisfactory results except for a gate, which makes quiet audio quieter.
This completely suppressed the clicks. However, when applied to the entire episode,
it also distorted the start of a few sentences which began with single short syllables.
The A gate filter has a number of parameters which could be adjusted to try to deal with these
cases, although I did not spend the time to do so. Another solution to this distortion problem
is to simply not apply the filter to those parts of the audio, which are affected.
If we record the audio as a series of small individual files, it would be relatively easy to
filter before concatenating the files together, while skipping those files which contained audio,
which is not suited to this method. Here are the results of the experiments.
FFMPEG, AFFT, DN. This reduces the size of the ticks, but they are still present. However,
they may be reduced to a level, which is considered acceptable. FFMPEG, A gate. This was very effective
in removing clicks with the right parameters. However, it can introduce some voice distortion
in the form of cutting out the start of a few sentences which began with single short syllables.
This can be corrected with a very short attack parameter to turn off the filter when a detects
sound above a set threshold. FFMPEG, AR and DN. This was relatively ineffective.
Socks, noise, AR, ED. This was effective in removing the sounds between phrases. However,
it introduces some distortion in the voice signal. I also tried combining filters.
FFMPEG, AFFT, DN, followed by A gate. This combined two different filters.
First, used AFFT DN to suppress the background noise to a lower level without much voice distortion.
Then it applied the A gate filter to suppress the noise levels between words still further.
This got rid of virtually all of the background noise between words.
Here is a short audio sample from HPR 4637. First is the unfiltered audio.
Second is the filtered audio using the combined AFFT DN plus A gate filters.
Since the clicks are very quiet, you may not hear them unless you are in a quiet environment.
Quite a few listeners would probably not be aware of the perceived audio problem in this episode
if it had not been discussed here. Nonetheless, it makes for an interesting experiment. Here it is.
But many desktops are not. If nothing is output, but many desktops are not.
If nothing is output, overall conclusion for noise ticks. The AFFT DN combined with A gate filters
seemed to offer the best overall results when used with the right parameters.
However, the author Vance speaks very clearly and evenly and so his voice is ideally suited for
use with this filter. Another author's voice may not be as suited to this filter.
These socks noise RED profile-based filter offers various degrees of trade-off between
suppressing noise and distorting the voice signal. As to whether this is an acceptable trade-off,
depends on the particular voice in question and how easily understood it is under normal circumstances
without additional distortion. The AFFT DN filter may be a fairly safe filter to use on its own
while producing acceptable if not perfect output. Overall conclusions. I have presented only a
few of the experiments that I conducted. By overall conclusion after all of this
is that there is no universal audio filtering method that works best in all circumstances.
There are instead a number of tools in the toolbox and picking the right one for the job
takes a bit of trial and error. However, if you have a repeatable recording environment,
then once you have decided what tool you need, use your creative script for it so that you
can have a repeatable processing set up. These conclusions apply to voice pod casting.
Music has a different set of criteria and techniques that work well with basic voice pod
casting may produce poor results when applied to music, which has a broader range of frequency
and just as importantly a broad range of loudness. If you are used using filters and effects
in audacity, many of the settings on those correspond to arguments in the command line version
of FFMPEG. It is worth learning how to use FFMPEG directly to automate your recording process.
The experiments that I conducted were greatly assisted by writing scripts, which created
multiple versions of audio files with different settings, thereby allowing me to try many
different alternatives relatively easily. It also allowed me to concatenate different audio samples
into a single audio file and so listen to different versions in quick succession,
making subjective listening judgments more reliable. It is important to keep in mind in all this,
but I am playing with audio filtering, mainly to have fun. It is not necessary to do any of
this if you think your podcast episode sounds just fine without it. So don't let any of what
I have talked about in all this discourage you from simply recording a podcast and sending it in as
is. I will include copies of the filters I have described here in the show notes. Related matters.
Hardware characterization using audio signals. I found it useful to characterize the hardware
that I had in order to understand its limitations better before starting the experiments.
This is involved playing a signal out through a set of speakers and then recording it through a microphone.
I use two types of signal for this. One is a type of signal known as a chirp signal.
This is a sine wave that steadily increases in frequency as it sweeps across the audio spectrum.
The standard audio range is 20 Hz to 20 kHz, but for my purposes I limited the upper frequency
to 15 kHz to save time, as anything beyond that is not very useful for voice podcasts.
By recording the chirps signal with a microphone and analyzing it with a Fourier transform,
I could quickly see what each device was capable of.
See my previous series on simple podcasting for an explanation of what a Fourier transform is
and what software to use to see the results of it. Here is the chirps signal.
In addition to a chirps signal, I also use these series of simple tones of specific frequencies.
By using these tones of known frequency, I could gain an understanding of the limitations
of my speakers and headphones and just as importantly my own ears.
By understanding these limitations, I was able to narrow the range of frequencies that I need to
deal with quite considerably and set the high and low-pass filters accordingly.
These tones are a series of flag files generated with FFMPEG.
Here is a sample audio tone at 2 kHz frequency.
Copies of the script to create the chirps signal and the tones are in the show notes.
A not a review of some of the hardware that I used.
I said that I would not do a review of the hardware that I used.
However, some of it deserves mention for either how good or bad it was.
I will report each section using the hardware being described.
Maxwell headset. This is my original recording hardware. This is a headset with boom mic and USB
connection. There is no model number on it so I don't know the model.
This probably costs between $10 and $25. The earpiece is set on the ears and do not fully
enclose them. This makes it lightweight and comfortable to wear for extended periods of time.
It has a problem, however, with electronic noise consisting of a noise spike everyone
killer hurts. I was able to fix this with a series of filters using FFMPEG.
Fixing this problem is what got me started in understanding audio.
I will probably continue to use this headset to make podcasts.
Extra Headset. Model GH510. This is also a headset with boom mic and USB connection.
I purchased this headset for the purpose of experimentation for this podcast episode.
It cost $12.88. I've owned it to be surprisingly good for surprisingly little money.
It is fully enclosed earpiece, however, which may make it uncomfortable to wear in hot weather.
I may try doing some of my future broadcasting using this headset.
Born earpiece and in line microphone. This is a set of earplugs that go in your ears
and connected by wires and a very small microphone built into a small bolt in the cable.
It could actually be using a 3.5mm jack. The model number seems to be BUD250BL.
It costs approximately $3. I bought several sets of these and use them for listening to podcasts
when MP3 player. The earpieces are pretty good for listening with. A microphone works reasonably
well when used in a quiet location. It is less good when in a noisy environment.
It is very important, however, to secure the microphone to your lapel or other location
reasonably near your mouth and to point the microphone that is the small hole outwards
and not let it simply dangle freely. If you let it just hang, you will get poor quality
and inconsistent audio. Jan May condenser microphone model SF-910. I purchased this microphone
for the purposes of experimentation for this podcast episode. It cost $3.88.
As a condenser microphone, it is prone to picking up background noise more
and as such is probably not a good choice for podcasting by a single person sitting at a desk.
However, it is nonetheless a surprisingly good microphone for surprisingly little money.
I can't USB microphone model M306. I purchased this microphone for the purposes of experimentation
for this podcast episode. This has a USB connection. This is also relatively inexpensive at
$7.99 or roughly twice the price of the N-May microphone. Unlike the N-May, however,
it is absolutely rich. There was such a high degree of distortion when recording through it
that I found it could not use it in the fan experiment which I had bought it for.
I ended up buying the N-May microphone for that instead.
Easy effects software. The techniques described so far all involve recording audio files
and then processing them later to produce the desired result. This is probably the simplest
and most straightforward way of doing things if you are making a typical podcast.
However, there may be instances where you wish to apply filtering or other effects
on the live signal immediately and not after the fact. There is audio software which can hook
into your computer's audio system and do this with a live signal. For Linux, there is a package
called Easy Effects. This is free software and comes under a GPL V3 or later license.
I installed it from the Debian repository under Ubuntu 2404. You can create various filters
and even chain them together to combine them. I played with it a bit but do not know enough about
it to discuss it seriously at this time. However, I thought it would be worth mentioning
for the sake of those who bear wish to try it out themselves. Episode conclusion.
After having had some fun with audio and listening to other HPR members talk about audio,
I thought it would have some more fun by playing with noise reduction filters.
I have no intention of becoming an audio professional, but by doing some experiments, I learned a
few things and had some fun doing it. I hope that the rest of you found this of interest as well.
I will see you all again later in another episode of Hacker Public Radio.
You have been listening to the Hacker Public Radio podcast, at hackerpublicradio.org.
Today's show was contributed by a HPR listener like yourself.
If you ever thought of recording a podcast, then visit the HPR site to find out how easy it really is.
Hosting for HPR has been kindly provided by anhonesthost.com, the Internet Archive, rsync.net, and the HPR Community Content Delivery Network.
Unless otherwise stated, today's show is released under a Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.
