Is frequency analysis a viable attack on non-text data encoded by substitution (eg image and audio formats encoded with a substitution cipher)?
If one can prove that a large amount of image and audio data doesn't exhibit a frequency patter, only then can we consider frequency analysis as a non-viable attack.
For a modified simple example, let us say each guitar chord is encoded as a byte of audio data. If you analyse about 70+ songs, you will see that 4 chords are the most frequently used (as suggested by Axis of Awesome). Hence, a frequency analysis attack on each byte can be done for such musical data.
On an ending note, we humans are not good at generating data with uniformly distributed frequencies. So intuitively I feel that frequency analysis would be a possible attack for any data set somehow generated by humans (say playing a musical instrument). Is it computationally efficient? That is another question. But a formal prove and analysis of large data is needed either against or for this intuition to answer your question in a precise manner.