Extracted from taiwancards, where the pronunciation scorer builds
per-syllable acoustic prototypes from recordings that differ in speaking
rate. The algorithms are domain-independent, so they ship separately.

Pure Ruby, no runtime dependencies, deterministic. Median aggregation is
the DBA default: a 50% breakdown point keeps corrupt observations from
displacing the prototype. RBS signatures.