5 ms·
Strange that they are feeding raw audio in. Even in humans, there is a hardware transform to the frequency domain (the cochlea) before data is fed to the brain,
by briansm 4mo ago
Strange that they are feeding raw audio in. Even in humans, there is a hardware transform to the frequency domain (the cochlea) before data is fed to the brain, effectively doing this part in the LLM seems inefficient.
- nialse 4mo agoThe FFT is essentially just a matrix multiplication, or two. No need for fancy conversions. Just a huge amount of training data and a very large array.