parameter tips for talk-llama

[javadan]

Hey GG, great repos.

Anyone have any tips to improve talk-llama performance?

./talk-llama -mw ./models/ggml-base.en.bin -ml ../alpaca.cpp/ggml-alpaca-7b-q4.bin -p "George" -t 8

Not sure if I can tweak some parameters below, to get a more natural conversation going?

params.n_threads     
params.voice_ms      
params.capture_id    
params.max_tokens    
params.audio_ctx     
params.vad_thold     
params.freq_thold    

It works best if I talk as soon as it finishes talking with speak.sh, but it's not a guarantee.

Sometimes it just sort of makes up what i said. Usually some short sentence, like 'Thank you' when I said something else entirely.

[scalar27]

For me, that happens when there is background noise. If I use headphone or earbuds with a mic, it works much better. I would also like to learn what those parameters do.