1

Various initiatives, such as the European SESAR 3 program, the Airbus Speech Recognition Challenge, or Appareo's technology, highlight a growing interest in tailored Automatic Speech Recognition (ASR) solutions for air traffic control communications.

While my knowledge is limited to projects mentioned, I wonder about the broader implications and potential beneficiaries of a successful implementation of such technology. How could air traffic controllers, pilots, or airlines find value in this specialized ASR technology?

The primary application seems evident - reducing the workload for pilots and controllers through efficient transcription. However, as an aviation newcomer, I’m curious about the additional structured information that could be gleaned from transcribing and language-processing the ATC utterances - beyond what's currently available or inferable through structured sources like ADS-B or ACARS data.

A R
  • 11
  • 1
  • Welcome to Aviation.SE. Are you associated with any of the sites that you linked to? – Ralph J Feb 04 '24 at 01:33
  • 2
    If my experience with Chat GPT is any measure, AI has a LONG way to go before it can be entrusted with peoples' lives. – John K Feb 04 '24 at 01:36
  • @RalphJ Thanks! I'm not associated in any ways with the linked projects above. I'm mainly trying to understand if any sort of valuable piece of information could be extracted from transcribed ATC utterances - especially in non-safety critical situations. For instance, could relevant data be caught from an ATC transcription, transferred to ATM softwares, and enhance airlines' operations management ? – A R Feb 04 '24 at 01:57

2 Answers2

5

The difference in impact between a really, really amazingly excellent model, and a fairly good model, is worlds apart. If I don't understand what was said on the radio, I can ask the sender to "say again", and that works. If the model doesn't understand & guesses wrong, all sorts of bad things can happen. If it's frequently marking *** (i.e. unintelligible) in a transcript, that becomes a nuisance & a time drain for people who can do better without the extra layer between the speaker and the hearer.

Nobody cares if a YouTube video gets a transcription error; with ATC the stakes are exponentially higher. Until the accuracy is orders of magnitude better than we've seen so far, there are scant gains and significant risks from any such models. Eventually, with everything becoming good enough that the errors from the technology are less than the errors of what we do now, we'll probably see that sort of thing being adopted. How far out that point is, however, is somewhere between an opinion and a guess.

In the meantime, datalinks are gaining adoption (i.e. CPDLC) in certain phases of flight, and Digital ATIS (i.e. computer voice reading the text aloud, saving humans time in both the reading and the copying, since it's also sent over ACARS) have brought incremental gains. Speech to Text in any ATC context is probably pretty far off down the road, in my opinion.

Ralph J
  • 51,356
  • 17
  • 157
  • 249
  • 2
    Would suggest the key sticking point is that a carefully ATC optimised AI will fail at the worst possible time, since an emergency will by definition involve things outside the training data set. Human pilots and ATC can be trained for emergencies by being asked 'what would you do' questions. Current AI training would instead require a massive library of synthetic emergency communications being generated and then curated/annotated by experts. AI would also need to process out of channel context - eg engine failure on a single engine aircraft vs multi engine. – GremlinWranger Feb 04 '24 at 02:23
  • 1
    I agree with this assessment, we had a training system in Australia that was meant to work based on voice. It could handle the basic commands with no issues, but when it went to anything but the most basic of commands it failed to comply. We still use people to pretend to be pilots on that simulator even though it has this capability. – Bullfrog Feb 12 '24 at 09:19
0

I don't think is going to have a good impact implementing something like that, at least with the current technology. As others have mentioned, the risks can outweigh the benefits.

However, I do support 100% the opposite implementation of that, Text to Speech.

Not only I think the benefits would outweigh the risk, but also the suggestion of that implementation is to only use it for predefined messages so there is no guessing or any AI in the middle of the transmission.

Gabe
  • 2,353
  • 10
  • 34