Large language models are designed to generate text based on previous text. Translation from audio to text can be done via a neural net but it isn’t a Large Language Model.

Now, could you combine the two to say reduce error on words that were mumbled by having a generative model predict the words that would fit better in that unclear sentence. However you could likely get away with a much smaller and faster net than an LLM in fact you might be able to get away with using plain-Jane markov chains, no machine learning necessary.

Point is that there is a difference between LLMs and other neural nets that produce text.

In the case of audio to text translation, using an LLM would be very inefficient and slow (possibly to the point it isn’t able to keep up with the audio at all), and using a very basic text generation net or even just a probabilistic algorithm would likely do the job just fine.