Board index FlightGear Support Tools ATC-pie

[Support] Noises on FGCom; Voice recognition help

ATC-pie is a radar air traffic control simulation program for the FlightGear multi-player network.

[Support] Noises on FGCom; Voice recognition help

Postby nocusuzo » Fri May 25, 2018 11:37 pm

While using the Solo Game voice recognition I noticed that it almost always was picking things wrong.

While doing the FGCom echo test my audio comes with noise, similar to old extraterrestrial/alien movies.

Strangely, while doing the echo test, if I make some noise, for example, hit my desk, the noises stop!
Also it seems that if I start my PC and run the echo test it works fine but all other times the noises appear, even if I close and reopen ATC-pie.

This was also being reported here viewtopic.php?f=32&t=14181

I'm thinking in maybe using Jack instead of Pulseaudio and see if it would solve the problem, but I never used Jack before and don't know if it would help with this problem.


My specs:

ATC-pie 1.4.3
Debian stable + some testing stuff like FlightGear
Python 3.6.5rc1
Installed all python and libraries required for additional ATC-pie features
Last edited by nocusuzo on Sun May 27, 2018 9:39 am, edited 1 time in total.
nocusuzo
 
Posts: 12
Joined: Fri May 25, 2018 8:17 pm

Solo voice recognition. Was: [BUG] Noises on FGCom

Postby mickybadia » Sat May 26, 2018 7:37 am

Hi,

nocusuzo wrote in Fri May 25, 2018 11:37 pm:While using the Solo Game voice recognition I noticed that it almost always was picking things wrong. While doing the FGCom echo test my audio comes with noise, similar to old extraterrestrial/alien movies.

NOTICE: FGCom has nothing to do with solo. It is only controlled by the radio dock, i.e. usable in teaching or multi-player sessions. Now as I noticed the double-post and FGCom being a separate program, I will keep this answer about the voice recognition issue. For the FGCom issue: fu2 2nd post.

What does "picking things wrong" mean? Does it understand something different to what you are saying or do you get the unrecognised voice instruction notification (low beep if sound is on)? In the former case, what pattern do you see? Note that to better diagnose this you can use the solo cheat: "show recognised voice strings in status bar" and see what is actually being picked up.

In any case, what is your accent in English? If not native American, you will be departing from what the default acoustic model is expecting. I personally get things to work with a British accent, but sometimes need to force a few "rrr" sounds or distort my input with other American features to help out.
mickybadia
 
Posts: 397
Joined: Tue Sep 24, 2013 9:12 am

Re: [BUG] Noises on FGCom

Postby wkitty42 » Sat May 26, 2018 4:44 pm

nocusuzo wrote in Fri May 25, 2018 11:37 pm:While doing the FGCom echo test my audio comes with noise, similar to old extraterrestrial/alien movies.

if i'm understanding you, it is supposed to do that... FGCom tries to accurately simulate real radio comms... that includes static and similar noises...
"You get more air close to the ground," said Angalo. "I read that in a book. You get lots of air low down, and not much when you go up."
"Why not?" said Gurder.
"Dunno. It's frightened of heights, I guess."
User avatar
wkitty42
 
Posts: 5018
Joined: Fri Feb 20, 2015 3:46 pm
Location: central NC, USA
Callsign: wk42
Version: git next
OS: Kubuntu 14.04.5

Re: [BUG] Noises on FGCom

Postby mickybadia » Sat May 26, 2018 7:39 pm

Indeed FGCom has a few filters that make it sound more like real radio.

What I am saying is or implies the following:
* FGCom echo test without headphones will typically sound that loud obnoxious and repeating noise (resonance)---though you should be able to hear something of your voice back over that noise if you speak loud and clear enough
* if you want to echo test and hear something clear back, try again but with headphones
* resonance is NOT a problem outside of echo test (even without headphones) because mic is muted until PTT, and PTT deafens speakers
mickybadia
 
Posts: 397
Joined: Tue Sep 24, 2013 9:12 am

Re: [Support] Noises on FGCom; Voice recognition help

Postby nocusuzo » Sat May 26, 2018 10:07 pm

I'm using headphones while doing the echo test.

After tweaking my microphone volume it seems a little better, but the noise still appears.

-------------------------------------------

English is not my native language, maybe this is affecting the voice recognition, but I try to speak as clear as possible and I recorded my voice and it seems OK.

After tweaking my microphone volume it seems that the voice recognition is a little better, but I still have problem while trying to identify an aircraft by callsign. Usually I spell all callsign letters and numbers, for example, KQA7473 is Kilo Quebec Alpha Seven Four Seven Three, but the program understands it as 2-Z-Q-G-F-7-4-7-3 or K-Q-G-5-7-4-8-7-3 or K-Q-G-4-7-4-7-3 or K-Q-G-F-7-4-8-7-3

The same happens while identifying things for route like "proceed direct PHNL" proceed direct Papa Hotel November Lima, it usually gets the character spelling wrong, so I have to repeat many times until it works.

Am I doing it wrong by spelling all characters in a callsign?

Usually when I abbreviate the callsign it works, for example Kilo Quebec Seven Three worked.
Also using the airline callsign, for example, Speedbird One Seven Four Two sometimes works better than spelling characters.


It seems to understand fluid things, for example "cleared for takeoff" or "cleared for ILS".


While pronouncing the characters should I do it fast or stop after each character?

Is there any place I could tweak the voice recognition?
Last edited by nocusuzo on Sun May 27, 2018 9:41 am, edited 1 time in total.
nocusuzo
 
Posts: 12
Joined: Fri May 25, 2018 8:17 pm

Solo voice recognition. Was: [BUG] Noises on FGCom

Postby mickybadia » Sun May 27, 2018 7:37 am

Please read the following as they contain most of your answers:
  • "quick reference" (F1) tab on solo voice instructions
  • the resources/speech/Notice file
As you will read there, it is a good idea to try the loudest microphone level you can get without distortion.

nocusuzo wrote in Sat May 26, 2018 10:07 pm:Usually I spell all callsign letters and numbers, for example, KQA7473 is Kilo Quebec Alpha Seven Four Seven Three, but [...]
Am I doing it wrong by spelling all characters in a callsign?

Yes, callsigns are not meant to be spelt out. The program is recognising the best possible match within expected phraseology, which for callsigns is either an airline callsign and a flight number, or a tail number. Your example should be called "Kenya seven four seven three", or "Kenya seventy-four seventy-three".
I guess I could add the possibility of recognising commercial flight callsigns from spelt-out names, as it is not standard but sometimes (though rarely) used as fallback to work around confusion. But it is not yet implemented, so will currently fail if you do so.

nocusuzo wrote in Sat May 26, 2018 10:07 pm:It seems to understand fluid things, for example "cleared for takeoff" or "cleared for ILS".
While pronouncing the characters should I do it fast or stop after each character?

Clear-cut answer, quoting the quick ref: "Do not break, backtrack or silence too much between your words. Try to have your message flow from start to end."

nocusuzo wrote in Sat May 26, 2018 10:07 pm:Is there any place I could tweak the voice recognition?

Yes, you can in theory adjust it entirely to your pronunciation, but be prepared for a little work if you try. As I did not like to serve only Americans (it is only I don't like, not Americans lol), I implemented an option to provide a custom acoustic model for recognition, see solo system settings. If you are ridiculously lucky, you find one for "English with your accent" (or even yet: "... for aviation phraseology") on the internet. But since this is unlikely, the solution is to train your own model. At this point I have to let you read and follow the PocketSphinx documentation on training acoustic models, but the bottom line is that you need a relevant (thus significantly long) set of instructions for which you record your voice.
So it is up to you, but if you ever do that, please share!
mickybadia
 
Posts: 397
Joined: Tue Sep 24, 2013 9:12 am

Re: [Support] Noises on FGCom; Voice recognition help

Postby wkitty42 » Sun May 27, 2018 9:45 am

nocusuzo wrote in Sat May 26, 2018 10:07 pm:Usually I spell all callsign letters and numbers, for example, KQA7473 is Kilo Quebec Alpha Seven Four Seven Three, but the program understands it as 2-Z-Q-G-F-7-4-7-3 or K-Q-G-5-7-4-8-7-3 or K-Q-G-4-7-4-7-3 or K-Q-G-F-7-4-8-7-3

yeah, don't do that... it doesn't know that "k" is "kilo"... just say it like it is written... kqa7473...
"You get more air close to the ground," said Angalo. "I read that in a book. You get lots of air low down, and not much when you go up."
"Why not?" said Gurder.
"Dunno. It's frightened of heights, I guess."
User avatar
wkitty42
 
Posts: 5018
Joined: Fri Feb 20, 2015 3:46 pm
Location: central NC, USA
Callsign: wk42
Version: git next
OS: Kubuntu 14.04.5

Re: Solo voice recognition

Postby nocusuzo » Sun May 27, 2018 10:22 am

mickybadia wrote in Sun May 27, 2018 7:37 am:Please read the following as they contain most of your answers:
  • "quick reference" (F1) tab on solo voice instructions
  • the resources/speech/Notice file


I had read the "quick reference" (F1) tab before, but since it was having this issue I tried to have a small stop between letters/numbers spelling.

I just read the resources/speech/Notice file

mickybadia wrote in Sun May 27, 2018 7:37 am:[...]callsigns are not meant to be spelt out.[...]


Now it seems clear why the program was picking it wrong since I'm not supposed to spell the letters individually.

mickybadia wrote in Sun May 27, 2018 7:37 am:[...]I guess I could add the possibility of recognising commercial flight callsigns from spelt-out names, as it is not standard but sometimes (though rarely) used as fallback to work around confusion. But it is not yet implemented, so will currently fail if you do so.


It would be cool to have this feature 8)

mickybadia wrote in Sun May 27, 2018 7:37 am:[...]I implemented an option to provide a custom acoustic model for recognition, see solo system settings[...]


I didn't notice it was right there on the end of System > System set up > Solo session tab
I will check for other acoustic models available, and maybe create my own, who knows :D

-----------------

Since this is not a bug I changed the topic title.
nocusuzo
 
Posts: 12
Joined: Fri May 25, 2018 8:17 pm

Re: Solo voice recognition

Postby mickybadia » Sun May 27, 2018 1:03 pm

mickybadia wrote in Sun May 27, 2018 7:37 am:I guess I could add the possibility of recognising commercial flight callsigns from spelt-out names

Actually reviewing the code, it seems it is already doing that. But obviously prone to more recognition errors than an airline callsign token making things much less ambiguous... and a million times more realistic. Of course you need to know what they are.

nocusuzo wrote in Sun May 27, 2018 10:22 am:I will check for other acoustic models available, and maybe create my own, who knows :D

Do tell if you find anything. Many are or will be facing the same problem.

wkitty42 wrote in Sun May 27, 2018 9:45 am:it doesn't know that "k" is "kilo"... just say it like it is written... kqa7473...

Negative: ATC-pie knows that "K" is "kilo", and would in fact not recognise "kay" as "K". I will be very surprised if you tell us you had it work while spelling out callsigns reading English-named letters. Now as I said his "KQA" should be read "Kenya" anyway :-)
mickybadia
 
Posts: 397
Joined: Tue Sep 24, 2013 9:12 am

Re: Solo voice recognition

Postby nocusuzo » Sun May 27, 2018 1:58 pm

How would I pronounce the Callsign of a small civil aircraft, like JA-ITJ when using the voice recognition?
nocusuzo
 
Posts: 12
Joined: Fri May 25, 2018 8:17 pm

Re: Solo voice recognition

Postby mickybadia » Sun May 27, 2018 2:29 pm

nocusuzo wrote in Sun May 27, 2018 1:58 pm:How would I pronounce the Callsign of a small civil aircraft, like JA-ITJ when using the voice recognition?

Side note: this is a rather fictitious callsign as Japanese registration does not include the dash and uses numbers.

Regardless, as stated in the reference: "Abbreviated tail numbers for non-commercial flights are OK as long as at they end with at least two of the last letters or numbers." So this aircraft could be called "juliet alpha india tango juliet", or "india tango juliet", or simply "tango juliet" if there are no homonyms.

Dropping only one letter would be strange, but as in real life, callsigns like say G-ABCD can (indeed should!) be abbreviated "golf charlie delta".
mickybadia
 
Posts: 397
Joined: Tue Sep 24, 2013 9:12 am

Re: Solo voice recognition

Postby wkitty42 » Sun May 27, 2018 3:34 pm

mickybadia wrote in Sun May 27, 2018 1:03 pm:
wkitty42 wrote in Sun May 27, 2018 9:45 am:it doesn't know that "k" is "kilo"... just say it like it is written... kqa7473...

Negative: ATC-pie knows that "K" is "kilo", and would in fact not recognise "kay" as "K". I will be very surprised if you tell us you had it work while spelling out callsigns reading English-named letters.

i never said i did anything with it... i was going by what had been written previously and (mis)understanding that Kilo Quebec Alpha was incorrect... especially after seeing the following...
mickybadia wrote:[...]callsigns are not meant to be spelt out.[...]


mickybadia wrote in Sun May 27, 2018 1:03 pm:Now as I said his "KQA" should be read "Kenya" anyway :-)

so this is a reversal of the above quote, then...
"You get more air close to the ground," said Angalo. "I read that in a book. You get lots of air low down, and not much when you go up."
"Why not?" said Gurder.
"Dunno. It's frightened of heights, I guess."
User avatar
wkitty42
 
Posts: 5018
Joined: Fri Feb 20, 2015 3:46 pm
Location: central NC, USA
Callsign: wk42
Version: git next
OS: Kubuntu 14.04.5

Re: [Support] Noises on FGCom; Voice recognition help

Postby nocusuzo » Sun May 27, 2018 9:58 pm

I have read some things about CMUsphinx and on this page: https://cmusphinx.github.io/wiki/faq/#q-what-is-sample-rate-and-how-does-it-affect-accuracy it says the acoustic models available are based on 16khz and 8khz bandwidth, so I edited my Alsamixer config to use 16khz and downloaded and set to use the 16khz model, it seems the voice recognition is a little better than before now but I still have trouble having it understanding stuff, especially spelling things.

-----------

Also, something strange happened on one of my voice commands. For an airplane XV-ORP I said Oscar Romeo Juliet Runway 08 Right Cleared for take off, ATCpie understood it and wrote down XV-ORP Runway 08R, cleared for take-off, wind 030°, 18 kt and the airplane responded: Unable. Ready for departure from 08R. Wrong runway? shouldn't the plane take off since the runway I said was cleared for take off is the same it's waiting on?
nocusuzo
 
Posts: 12
Joined: Fri May 25, 2018 8:17 pm


Return to ATC-pie

Who is online

Users browsing this forum: No registered users and 1 guest