AI & LLM Features

Experimental voice recognition and Gemini phrase variation

NextATC currently uses AI in two separate areas: voice input and optional dialog variation. Both are experimental. They are designed to make the radio feel less repetitive and more natural, not to replace the normal ATC logic.

The important rule is simple: NextATC still decides the clearance, runway, altitude, frequency, route, taxi instruction, traffic state, and flight phase. AI may help with recognising what you said, or with rephrasing an already prepared ATC message. It is not a free-form chatbot and it is not an autonomous controller.

Experimental status#

The related Options entries are marked as experimental:

AI Push-to-Talk Recognition (Experimental) for voice input.
Dialog Generation and A.I. (Experimental) for generated ATC wording.
AI - Gemini (Cloud, Experimental) for cloud LLM phrase variation.

Experimental means that results can vary by microphone, accent, aircraft noise, internet connection, model availability, API limits, and the exact message being generated. If something sounds wrong, too slow, or unreliable, turn the feature off and use the Radio Log buttons. The flight continues with the standard built-in phraseology.

What NextATC uses AI for today#

NextATC has two AI-assisted features:

Push-to-Talk Recognition listens to your microphone and tries to turn your spoken readback or radio action into text.
Dialog Generation and A.I. asks an LLM to rewrite some ATC messages so they do not always sound identical.

These systems are independent. You can use voice recognition without Gemini, Gemini without voice recognition, both together, or neither.

Push-to-Talk recognition#

Push-to-Talk recognition is speech-to-text, not an LLM conversation. You bind the NextATC/voice_push_to_talk command, press or toggle it, and speak the readback or current radio choice. NextATC transcribes the audio, shows a preview in the Radio Log, and checks whether the important operational content matches what ATC is waiting for.

For example, callsign, runway, squawk, heading, altitude, frequency, and procedure names matter. You do not need to speak every word exactly like the button label, but the key numbers and names must be present and recognisable.

There are two operating modes:

Hold-to-Talk: hold the PTT key while speaking, then release it. This is the default and closest to a real radio.
Toggle Talk: press once to start listening and again to stop. This is useful when the assigned button is awkward to hold. Guided auto-toggle can also start and finish listening around supported prompts.

Push-to-Talk models#

The Options window lists several recognition models. Treat Whisper Base English as the practical baseline today.

Whisper Base English (about 142 MB): the stable embedded model and the recommended starting point. It is English-focused and usually the best choice for standard ATC readbacks.
Whisper Small English (about 466 MB): a larger embedded Whisper model that can improve recognition at the cost of more disk space and slower transcription.
Whisper Medium English (about 1.46 GB): the highest-accuracy embedded option currently offered for consumer PCs, intended for faster systems.

Only these embedded Whisper model sizes are offered for local transcription. English is the default. Multilingual model support switches the selected size to a separately downloaded multilingual Whisper variant and lets Whisper detect the spoken language automatically. Use it for local, non-English proper names or pronunciation, not to change the ATC interaction: the readback text that NextATC expects remains English. If a larger model is too slow or less reliable on your system, switch back to Whisper Base English.

Gemini dialog generation#

Gemini is used only for text generation. It does not listen to your microphone, and it does not decide what clearance or instruction you receive. When AI - Gemini is selected and an API key is configured, NextATC sends a short text request containing the message to rephrase and the aviation details needed to preserve the same facts. The Model button in Options lets you choose Gemini Flash or Gemini Flash-Lite; NextATC always uses the latest version of the selected model published by Google.

Gemini then tries to create a different wording for the same ATC message. The goal is variation: different openings, slightly different cadence, and less repetition during long sessions. The route, runway, squawk, altitude, frequency, and other operational items must remain the same.

The built-in phrase is still the reference. If Gemini is disabled, unavailable, too slow, returns empty text, or produces a phrase that fails validation, NextATC falls back to the built-in message.

The Gemini variation button in Options cycles through four experimental inference styles. Use it when you want to compare how Gemini rewrites the same kind of ATC message: more conservative, more varied, more resistant to copying the template, or more guided by examples. It changes wording style only; it does not change ATC decisions.

Local LLM models#

Local LLM entries may appear in Options as AI - <model> (Local paused, Experimental). In normal builds they are listed but paused. They are not the recommended path for current users. Use the built-in NextATC mode for strict predictable phraseology, or Gemini if you want to test cloud phrase variation.

What to expect#

AI phrase generation can make the radio sound more alive, but it can also be conservative, subtle, or occasionally wrong. Even with validation, a generated line may sound unusual, omit nuance, or choose wording you do not like. If you want the most predictable experience, leave dialog generation on the built-in NextATC mode.

For voice recognition, expect best results with a headset microphone, a quiet cockpit, short readbacks, and clearly spoken digits. Always keep the Radio Log visible when testing. The mouse Readback and Say Again buttons are the reliable fallback.