As many know about me (or think they do), I have
always struggled to write without making typos. You might recall how my texts
used to look in the past. The main problem is my unruly fingers, which don’t
always land on the right keys—or they hit some key next to the intended one. It
has now been more than two and a half years since I started working with AIs,
beginning with a software solution for text correction.
I am here to describe what I have achieved with it,
not only in terms of text quality but also in other areas related (or not) to
DXing and with the help of software “robots” as many call it even if th truth
is AI are just software. I am also introducing you Quilbot a web program I use
long time for text editing. More below
Let me know if you'd like a version with slightly more
formal phrasing or one that keeps even more of the original sentence
fragmentation for character.
First let mε introduce you the programs I
use starting from my best :
DeepSeek
is my go-to for everything—completely free and packed with abilities,
especially in its latest update, which I use with the passion of a teenager.
Gemini
is the AI from Google.
Meta, I use
just for fun.
Claude, the
flagship from Anthropic, is very limited—only three dialogues every four hours,
suitable just for trial purposes. I don’t intend to subscribe unless the price
drops below $15/month.
Grok I use
for limited work, especially for making detailed documentation.
Perplexity is only for testing and some meta-analysis.
ChatGPT
was the first AI I used, but I don’t use it very often now, since my top
programs can do all the work for free.
That’s nearly
all the most known !
My current DXing-AI scenario is very complicated,
since it affects almost everything—much more than I could have imagined. In
addition to everything else, I still use Quillbot, a fantastic web app also
used by university students. It does more than just correct text; it acts as a
grammar and writing assistant.
I have been using it for a long time (since 2022?) to
replace Google Translate and DeepL, because it has many more capabilities for
refining and handling text. Its features outperform my standard way of working
by offering—even in the free mode—text restructuring, translation,
paraphrasing, humanizer, AI detector, summarizer, plagiarism detector, picture
maker [?], plus a set of PDF and image handling tools in their latest version .
Most of these features are in the premium
subscription. Don't forget: Quillbot's core is AI-based!
Here are the features I use AIs for today, and how
Quillbot can also help with most of them.
-Text correction ;
I first started using Quillbot back in ’22 to clean up
my typos. It helped a lot—caught about 90% of my mistakes. That said, I still
had to go back and fix a few words manually, leaving maybe a false 5% error
rate. That remaining 5% includes some original typos, mostly in my DX-related
text.These days, any decent AI can do better, hitting close
to 100% accuracy. But for DX purposes, sometimes you need to keep a certain raw
style to make the text punchier.
One thing I’ve noticed: Gemini is the most “agile”
when it comes to fixing typos—even when my writing is a total mess. Here’s how
Gemini replied when I asked if it could make sense of my typos.
Contextual
Analysis: I look at the entire conversation history to figure out what you
are talking about. Even if a single sentence is hard to read, the surrounding
text usually gives me the clues I need to understand your meaning.
Pattern
Recognition: I am trained on vast amounts of text, including common
misspellings, typos, and grammatical errors. This allows me to
"correct" what I'm reading in real-time to match the most likely
intended word or phrase.
|
Your bad typing
|
How I understand
|
|
Missing letters
|
Guess from word shape
|
|
Wrong order (nglsi)
|
Rearrange letters
|
|
Mixed Greek/English
|
Try both dictionaries
|
|
Repeated keys
|
Collapse them
|
|
Nearby keys (m instead of I)
|
Keyboard distance
|
Focusing on
Intent: Instead of just reading the literal letters, I prioritize
understanding what you are trying to achieve. If you ask a question, I look for
the core subject and the type of information you are seeking
#IN nearly the same way, the remaining AI programs can
understand someone's bad or garbled writing. Just see what DeepSeek replied.
-OCR :
This is now the most useful operation for most AIs. DeepSeek
is the best among them (and free) and can scan texts with 90% accuracy. I have
tested it with screenshots on my PC, but also from taking shots from newspapers
and handwritten text, with very low success for the handwritten text. Its
limits are 50 per chat. It is not good with handwritten text; read below for
more.
Gemini, Grok,
and ChatGPT are also OK, but with limits: just one picture per day for GPT
and Grok, Gemini with only 5 per chat bubble. Here are results
for Grok and ChatGPT as checked with Gemini.
For chat gpt The OCR accuracy is approximately 35%, as it
frequently failed to identify correct frequencies, times, and technical
terminology. Significant errors include misinterpreted station names, missing
signal data, and garbled text structures. This poor performance renders the
transcription unreliable for accurate radio logging, requiring a manual
re-entry of all data.
Grok :The OCR
accuracy for this text is approximately 25%. The transcription exhibits severe
degradation, including frequent character substitutions, garbled numbers, and
misinterpreted technical logs. Essential data, such as frequency values and
station identifiers, are consistently corrupted, rendering the log
unintelligible and unusable for accurate radio signal documentation or
historical analysis.
Deepseek
could not recognize handwriten text
Text Translation:
I used the same scheme as above. Originally I used
Google Translate with quite bad results. DeepL was quite better, even if the
wording was nearly the same. Both have some limits in their free operation:
- DeepL could translate only a few times per day, around
5.
- Google
Translate required texts
up to 5,000 characters.
- The
premium version of QuillBot does not have such problems, except in older times leaving some
parts of the original text inside the translated text.
Still see it having a limit of translation any text. As
for example a full issue of WORLDWIDE
DX CLUB Top News could not be translated
I have also seen some other bugs which have been
solved over time. For example, if a paragraph is very long, Quillbot showed an
error message that it couldn't handle the text. Texts with mixed language could
not be supported either. In a very long German text (>4k), I saw that parts
of the original text had been passed to the translation pane.
More recently, Google Translate could translate mixed
language (e.g., Greek and English) into Malaysian.
Any AI program can identify the original languages and translate a mixed language text into the requested language ie English
Summaries:
With the proper prompt, it can summarize your text
in the default language or even directly into another language. Many times,
while dealing with other languages, I could have direct summaries from the
original transcript texts, but especially Gemini thinks that it is
preferable to manage the texts into bulleted or numbered format. Sometimes
you can intervene in the result by using a more or less formal format, or by
adding the ’color’ if you wish (casual, satiric, etc.).
BTW, QuillBot has a separate summary feature with 4
levels, from short to very long summary, also using a keyword from a proposed
few to adapt the summary or even add some color as noted above.
-Pictures:
I am not very sure how picture generation can be
useful for any DXer, except for making a useful "business card"
as a gift from a radio station. I have been a sub in two specialized web apps
(Ideogram and Leonardo AI), two of the flagships in the industry. Leonardo AI
is more than a picture generator, It uses a big set of engines to support
impressive pictures and very short videos. Ideogram can make pictures and DTP
images.
Now most of the AI web apps can design pictures, starting from
GPT with their own model. Imagine is the name for Grok, and Nano
Banana for Gemini. I still think that these apps are a bit behind Ideogram
but even further behind the Leonardo environment.
BTW my older cards
made for the two radio programs
made for GMR and dangdut were made with a simple DTP program
-Transcription:
As you know, I have used various systems to transcribe
long ago since the "birth" of AI in the last years, which I prefer
not to disclose except for the newer Any2txt, which is the toughest model to
transcribe any recording for the better.
Now, continuing the information, most of today's AI
programs for transcription support limitless time for transcribing, such as Turboscribe
and Yescriber, while others like Any2txt are time dependent with
400 minutes for the basic set.
Most models support 3 levels of accuracy, and this
makes them search the higher the file. Based on Whisper engine, which was
released near the end of '23.
Google
was also in this wagon since those "old times by asking users to add Whisper into their
"drive" space and run a set of macros to make transcription easier. I
used it, but it's been a long time since I stopped… hacking! Every time it
required running a macro and then waiting a few minutes until the results
ended. Whisper was a little better, but it required clean audio in order to get
nearly error-free results. I know that a few DXers also tested using that.
More recently, offer direct transcription and dig more
into audio analysis. In the end you will
see two appendices. The first shows all its capabilities (A) as an DX assistant with many examples and (B)
as a transcription software cross tested with two more . The only things I can
note here are: it supports more languages than any other web app, such
as Mongol Tibetan or Uighur, for example.
Most audio and video formats are supported, with 5
files max per "chat bubble " with max time below 10 minutes, even if the system says it
can support up to 5 hours.
-Handling transcripts
(embedded text as in my DX logs)
This is something very easy to do without digging more
into the digest as shown below. It is something that is rather
"default" into Gemini or can be made with limited process in every AI
web app, except the design of the prompt.
Just first add all the original or translated
transcripts together with their date and time, and ask the app to make the
remainder. I prefer to use a 50-word summary for each log.
Next is the daily digest using the same concept. The
only difference is handling the full transcripts and an external program
(Simplenote)
(B) Making the day's digest from transcripts
As you all know, I many times prefer to collect all
the translations into a final digest posted into Simplenote.
The process I do is simple (old way):
- Copy-pasting
the translated texts from the web apps into a file (and via Google
Translate or QuillBot)
- Add
the filename as a header to show date, time, frequency, and most times the
WebSDR name
- Since
I use Simplenote, a point is needed to separate paragraphs; this is made
with a short macro in Word
- Copy-paste
all the text into Simplenote and a name into the header
Making things simpler nowadays , Claude does all the
process from start to the end, i.e., the "publishing." Sometimes it
goes even farther by adding texted lines. Just look below at the two modes:
http://simp.ly/p/DncW7l the first is
in digest
http://simp.ly/p/kV9sDj a bit raw
mode transcription made 3 weeks before.
I already have a
very good experienceih this program making vcey nice docx files
with rabbles ad patterns in cases of
more complex texs that required more pro layout.
-Text scraping
for offline use
A few times I have made a table based on the EIBI
list for a radio station for offline use (i.e., printed form). It was tested in
Gemini, DeepSeek, and ChatGPT. All are good in their work,
but DeepSeek seemed quite better in sorting the table; otherwise, it takes a
short time to format the text then table.
In a different way, these days Gemini provided
me in zero time a comparative list of supermarket prices between Greece and
another country.
Shazaming music / songs
This can happen in two cases:
Original Shazaming: Shazam identifies songs by creating a digital
fingerprint of an audio sample. It converts the sound into a spectrogram,
extracts key frequency peaks, and compares this unique pattern against a
massive database to find a matching audio track.
Gemini,
in its recent update [28/5], can also track songs with its mixed technology and
can identify the song quite easily, with song and singer. There are still
missing songs in its playlist, especially from the various songs in the inner
Chinese counties or in Myanmar, etc.
Link using and making summary
This was how I first started two years ago: I made
summaries related to ERT or other cases and posted them on radio forums,
including WOR. Any of these AI apps (software apps!) used the link I provided
and generated a proper article summary. Simplicity itself—just a short prompt..
- meta summary (summery of multiple AI
summaries)
I wanted to go deeper into the subject. Instead of
sticking to only one AI, I decided to use more than four. I would take all the
articles, combine them into a new summary, or even merge all opinions into one
long, complete story. In the end, one of the AIs would finalize the process.
For example, I would start with GPT, Gemini, and Claude, and then use DeepSeek
to summarize or blend all the texts. It is the only program that can handle
very long texts.
You already seen my
latest posts : The Eurovision asia is one of these cases. There are a few more article
sin my pages in both zliangaslogs page and my more general page
-Predefined scripts for tailor making other text
etc tasks
Sometimes it is better to let AI correct my texts—not
only for clarity but also to provide basic reasoning, acting as a lawyer might,
or even to rewrite the entire text for smoother English. As you know, I am not
a native English speaker. Therefore, AI can go beyond simple corrections and
dig deeper by rewording the text to achieve better English fluency.
- Identify the closest transmitters for a kSDR
Two times I used it together with the MW list table to
find the closest transmitter near to the kSDR used. The results were a bit
weird as I was still in doubt about the results.
- Preparing the email for a radio station for
requesting a QSL
This process began with an audio clip containing all
the relevant information for a radio station report. Gemini asked me if it can prepare
a QSL hunter email based on that clip.
Gemini handled everything: it searched for the
station's email address (though not always accurate), composed the email with
sender, receiver, and subject lines, formatted the text according to standard
QSL hunting practices, organized the monitoring details (date, time, signal
quality, SNR, etc.), added the program log, identified any songs present, and
arranged all the information into a professional-looking email.
Simpler text
handling not reoffered above
- A2z sorting if you don’t want to use word to do that
- text cleaning
I had some cluttered texts after
scanning a very large document with a manual OCR program. This simple option
not only cleaned up the garbage but also translated the text into Greek.
Something that seems not so easy: if you ask, for
example, to keep the dialogue in your original language while handling the rest
of the text in another language, it turns out to be quite difficult. You need
to be very persistent, guiding it with more instructions than expected.
Many times, these programs forget what you previously
wrote, especially from very old discussions in the same thread.
.........
The full text is
very long for here as it contains three more things:
Part A list of text handling by DeepSeek
Part B Gemini as your best
Dx assistant – a nice discussion with Gemini on what it can for a DXer
including some analysis with 🥝SDRs 😊
Part C transcription measurements between Ais. Which is the best?
You can fin that
as a PDF article
https://app.box.com/s/d3l4o9ib498easwdkbmdxl75wo1tpdfy
