I am working on fictional single words (names) generator that have to sound like words from a given sample. I have the generator up and running that gives reasonable words 70% of time. I thought of improving this value, ideally to ~99% (so no manual step is necessary to discard clearly missed words).
I've already attempted fuzzy matching to any word in the initial sample (and rejecting words below given thresholds), and this winded up reasonable answers to 95%-99%, but in my experiments this approach is rejecting too many well-sounding names (even 60-70 rejections per 1 generated name - this removes a lot of entropy from the initial set).
Now, I am thinking of using classification instead to tackle the problem from another angle: using classification to discard the generated name.
The problem I have is I have to vectorize a single word based on its sound for classification. I've already excessively searched the Internet and the only solution I see are for vectorizing texts, not single words.
So, the question here: what kind of algorithm for vectorization of single fictional word based on its sounding have I missed?