Atomic-phoneme confusion analysis

Perform confusion analysis at the atomic IPA-segment level for the phoneme-track tokenizers to determine whether the tokenizer-invariant error patterns observed at the whitespace-separated SAMPA-word level persist for individual phonemes.

Background

The paper compares 18 phoneme tokenizers and finds that their dominant substitution patterns are broadly similar when confusion counts are computed over whitespace-separated SAMPA words. However, this analysis does not resolve which individual phonetic segments are confused.

Because the phoneme-track systems emit phoneme sequences, atomic-segment analysis would provide a finer-grained test of whether the acoustic encoder or tokenizer determines the detailed error topology. The paper explicitly leaves this analysis unresolved.