[linrad] RE: Speech processing (Linrad-01.25)

[Leif Åsbrink <leif.asbrink@xxxxxxxxxxxxxxxx>]

Hi Jim,

> Over the years I have done quit a bit of experimentation with
> audio (and RF) processing for the SSB transmitter.  
My idea was to encourage a few people to do this kind of
experimentation under controlled circumstances so we would
know how large the improvements are (it is easy to listen
to the noise-free signal to hear how badly it sounds for
rag-chewing)

> It is my opinion that audio equalization is perhaps more 
> important then level processing.
Hmmm, I think the problem is a bit more complicated. 
If we limit ourselves to a filter - clipper - filter 
architecture I think you are right. The clipper is just
a clipper and it does not matter much how it is done or 
whether it is on AF or RF. The first filter has to limit
the passband to fit the second filter. Whether it also
gives different amplitudes to different parts of the 
passband is an open question to me.

> My conclusions are as follows:
> 
> 1) Start with equalization that reduces low frequencies and peaks
> audio in the 1 to 2 Khz range. This will vary depending on voice and mic.
The voicelab package allows you to set upper and lower frequency
as well as to set a slope in dB/kHz. I just played a little with
it, but I did not notice any significant improvement using a slope.

The corner frequencies of the filter preceeding the main clipper has
to be set for the same passband as the final filter. Running an 
audio clipper without a bandpass filter in front of it can not
replace a RF clipper.

> 2) The second stage should be a fast attack, slow decay audio
> compressor that will tend to keep the amount of processing constant.
Hmmm, do you mean that the bandwidth should be limited prior to this 
AGC?

One could use a wideband input, maybe evenm with some enhancement 
for high frequencies to make pulses really high. Clipping pulses
at high bandwidth before selecting the final passband (250 to 
2450Hz?) could be a good idea.

Before the audio compressor I think one has to filter out the 
desired passband.

> It should have a dynamic range of perhaps 10db. Noise gating 
> (-20db) can be used to limit background noise.
OK.

> 3) After the signal is converted to RF, a proper filter should limit the 
> bandwidth that enters the RF clipper.
The bandwidth can be set equally well at audio frequencies. On the RF
side one has to eliminate the other sideband to generate USB or LSB.
Operating on the DSB signal is equivalent to audio processing.

> 4) The RF clipper at this stage should be variable from about 6 db to
> 20 db of clipping.
OK. I would think it needs 0 to 30 dB or even more.

> 5) A matching RF filter cleans up distortion products. The filter can 
> have a slope to reduce some un-needed low frequency '
> energy. In other words it might be flat from 1Khz to 3 Khz, but 
> 500 hz may be 3 to 6 db down and 250 hz could be 6 to 12 db down.
Now, this is more interesting. Low frequencies may or may not have
ben suppressed already. At this point the signal is clipped in amplitude
so filtering it with a filter that has different amplitudes at different
frequencies might produce some interesting results.

Having a slope on this filter means that the clipping is incorrect. 
A sinewave at the low end of the passband would give maybe 10 dB 
less power than a sinewave at the high end. It can be managed by another
clipper or by another fast attack, slow decay compressor (an ALC) but
in order to not disturb other band users, one more RF filter is then
needed.


> This concentrates RF power in the frequency range that most of the
> intellegibility occurs. The filter could peak around 1.5 to 2 Khz and
> allow some additional rolloff at 3Khz, when 3Khz might otherwise 
> be considered
> the upper end of the filter (normal -3db point). This should 
> restore a more natural
> sound.
???????

> A full implementation of this might be difficult and expensive when
> done with hardware, but could be relatively cheap and easy when
> done in software. Some degree of varability may be desirable.
This is what the Linrad voicelab is all about. It implements the
things you suggest and many more things as well. The audio equalizer
is very primitive, it is just a single parameter setting the slope 
in dB/Hz. If this parameter seems to be useful at all, the next
step will be to allow a more flexible filter. If you are right in 
assuming that it is a good idea to suppress low frequencies I think 
I should get feedback saying a slope on the filter is a good idea. 
If people tell me they can not find that the slope is doing any good, 
then I tend to think other forms of low-pass filtering (6dB/octave 
below 500Hz or similar) would not be very useful.


> Most people would not have the ability to intelligently select various
> parameters, so it might be a good idea to have only 2 or 3 presets.
> When less processing is used, the bandwidth can be generally
> wider and when more processing is used, the bandwidth can be generally
> more narrow. When 3 settings are used it could be labled as follows:
> 
> 1) Ragchew ( little clipping and wide band ).
> 2) Contest ( moderate clipping and bandwidth ).
> 3) DX ( maximum clipping and minimum bandwidth for difficult conditions ).

> 
> As well, 2 or 3 frequency response presets for stage one could help
> tailor the rig to someones voice or choice of microphone. For instance,
> 3 switches could be selected on or off as follows:
> 
> 1) Bass cut on/off. ( -6db )
> 2) Midrange boost on/off. ( +6db )
> 3) Treble boost on/off. (+6db )
> 
> Hidden pots can be made available for the advanced user with
> default 6db settings for the beginner.
Linrad is available for advanced users only;-) It will come with 9 
settings that you can taylor as you want to have them. For 144 MHz
aurora it is very good to use 6 kHz bandwidth without any clipping
for example. (If your qso partner can use large bandwidths). It
is a way to get Q5 readability on voice in situations when normal
SSB is not copyable at all.

Unclipped narrow bandwidth or clipped at high bandwidth may be
advantageous if the interference is far from white noise.


> In my opinion:
> Audio clipping can give more intelligibility under weak signal conditions
> when the overall frequency response is not optimal. This is only due to
> the generation of harmonics that are heard over the noise better, 
> but harmonics
> are a poor substitute for those frequencies that are really 
> present in the human
> voice, thus the need for good equalization with RF clipping.
Well, this is theory. I am looking for experimental evidence:-)

Feedback from Pierre, ON4GN indicates problems so I will have to fix
those before I can expect any feed-back. Like everything else,
this
will take some time....


73

Leif / SM5BSZ


LINRADDARNIL