开云体育

On Sun, Oct 17, 2021 at 11:04 PM, Jerry Gaffke wrote:

Hans has a nice 50W class C PA for the QCX if you wanted

Hi Jerry,

That is a good idea. It is so easy to only think of transceivers now days. A separate transmitter and using the uBitx, or even a SDR, as the receiver is a good alternative.?

If I did modify the uBitx, I'd try to do it in such a way as to not mess up the bias for SSB, I'd use something as simple as a relay to switch the RF source and bias when trying the digital modes.?

I just made a trip to my museum of unfinished projects. Found a box with a really nice Turkish stamp of a very happy little girl operating a soldering iron on it, and found a 10W HF amplifier kit within. Along with some low pass filter kits. This could be a good start. I image I could convert it to class C without much trouble.?

Tom, wb6b

On Mon, Oct 18, 2021 at 05:18 AM, Evan Hand wrote:

I thought that FreeDV was a multiple carrier mode and would require a linear amp.

I was proposing to use the FreeDV codec but use the GMSK modulation type. Or another FSK type it it is better.?

I don't think there's any reason not to mix and match codecs (voice encoders) and modulation schemes.

Tom, wb6b

Gmsk is really very simple.

First, it picks two frequencies that the carrier switches between. Second, the frequency is switched preciselt where the signal hits the zero voltage. Thus, it avoids the abrupt ending of one frequency waveform and start of the next.

- f

The codec2 with packet overheads will come to approximately 1500 bps. With the 300 baud limit, we are looking at 1500/300 = 5 bits per baud.?

We slice this many ways. The easiest way is to have five tones. One for each bit. Further, you can reduce it to half, or lets say 3 bits with each bit having four values at 0, 90, 180 and 270 degree phase shift from the previous phase, thus conveying two bits. Further, you can add two amplitude levels double from four to eight different? values, bringing the bits per tone to eight.

These are usually plotted on a phase/amplitude graph as a constellation.

These higher bits per baud accordingly need better cnr to resolve the exact phase or amplitude.

Wspr/jt set of modulation schemes limit themselves to a single frequency modulated tone. It achieves remarkable relience by averaging the carrier frequency over tens of cycles

- f

Yes, that all sounds right.?

You can do a lot with 1 carrier, but the more you do the more
susceptible it is to noise.
I am curious how much more bandwidth 2 tone GMSK needs over something
using?
the most sophisticated multiple tones, multiple phases, multiple
amplitudes scheme
for equal performance in the presence of typical HF noise.?
I'd guess not much, but could be wrong.? Regardless, on HF we would be
forced
to do that stuff if we must follow this absurd max symbol rate of 300
baud.

D-Star on VHF gets by with choosing between just two tones, has a
bandwidth of 6khz
for fair voice quality.? Bitrate is higher than what Codec2 can do,
Codec2 should get it down to 3 or 4 khz for the same voice quality.

At 1500 bps, I'd guess Codec2 is just barely intelligible,
everybody sounding like a 50 year old TV robot.

Jerry, KE7ER

On Mon, Oct 18, 2021 at 10:58 AM, Ashhar Farhan wrote:

The codec2 with packet overheads will come to approximately 1500
bps. With the 300 baud limit, we are looking at 1500/300 = 5 bits
per baud.?
We slice this many ways. The easiest way is to have five tones.
One for each bit. Further, you can reduce it to half, or lets say
3 bits with each bit having four values at 0, 90, 180 and 270
degree phase shift from the previous phase, thus conveying two
bits. Further, you can add two amplitude levels double from four
to eight different? values, bringing the bits per tone to eight.
These are usually plotted on a phase/amplitude graph as a
constellation.
These higher bits per baud accordingly need better cnr to resolve
the exact phase or amplitude.
Wspr/jt set of modulation schemes limit themselves to a single
frequency modulated tone. It achieves remarkable relience by
averaging the carrier frequency over tens of cycles
- f

With 1.6 kbit/s you get very good voice with state-of-the-art AI-based
voice encoders (eg. LPCNet or SSMGAN). Please take some time to hear:


And try also with FreeDV 2020 (with LPCNet).

1.5kbit/s with audio barely intelligible was 25 years ago with some
second generation vocoders... which were indeed pretty good for the
time. Codec2 in standard version go down to 700 bps, and down to 450 bps
with data-driven reconstruction techniques.

Cheers,
Rafael

On Tue, Oct 19, 2021 at 04:34 AM, Rafael Diniz wrote:

With 1.6 kbit/s you get very good voice with state-of-the-art AI-based
voice encoders

Hi Rafael,

I see you have an interesting paper covering some experiments with the
new trained neural net voice encoders/encoders used on HF radio links.
'adio_de_longa_distancia_utilizando_a_banda_de_HF
<'adio_de_longa_distancia_utilizando_a_banda_de_HF>

For those interested, the pdf can be uploaded to google translate and
it will spit back a translated version.?

Tom, wb6b

Rafael,
Is a CW decoder possible with band waterfall in its input? I don't
know how to start with tensorflow. It would be dandy if someone can
write? a simple HelloWorld for us old timers to start using TensorFlow.
- f

On Thu, Oct 21, 2021, 5:09 AM Tom, wb6b <wb6b@...
<mailto:wb6b@...>> wrote:

On Wed, Oct 20, 2021 at 03:24 AM, Rafael Diniz wrote:

Also, for images, HIFIC, for example, provides significant better
compression than traditional non-AI based image encoders.

Hi Rafael,

That is really cool. I built an interface to TensorFlow many years
ago and had fun feeding images to a early neural net that was all
the rage at time. Haven't done much with AI in recent time. Looks
like some great strides have been made in real world applications.

Because you are involved currently in AI and Neural Nets, I have a
couple of random questions.

1) Regarding the AI image compression. Is it possible that there
would be adversarial patterns/images that could produce completely
unexpected results. An example was a university study where people
taped a piece of paper to their chest with a random looking
patterns printed on them. It fooled the AI into thinking the
pattern was a face and it spit out a "match" Was some time ago so
I don't remember exactly what the result was.

2) Something I've toying around with and trying to find any
possible related research. That is, would it be possible to train
an AI (possibility using the same training set as used for the
speech compression) to listen to the "Donald Duck" sounds of a SSB
transmission received by an AM receiver, to produce fully
intelligible speech? Reading about the AI speech compressors and
reading about the training data sets that are available, may make
it possible for me to give it a try.

I wonder how much it is going to drain my bank account setting up
an Amazon AWS server with GPUs and run the training?

Look forward to your paper.

Tom, wb6b