CrystalASR: Hierarchical Phoneme-Grounded Speech Decoding

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

We investigate a design principle for automaticspeech recognition where linguistic structure is explicitly enforcedthrough intermediate representations. The resulting system,CrystalASR, decomposes decoding into three modular layers:a 3.3M parameter phoneme CTC head, a zero-parameter rulebased word decoder, and an optional language model (LM) fordisambiguation. A defining constraint is strict upward information flow, ensuring higher layers modulate but never overridelower-level acoustic evidence.Experiments on LibriSpeech dev-clean show that CrystalASRachieves 17.44% WER while requiring 21× fewer trainableparameters and 14× faster inference than a comparable endto-end baseline. Error attribution reveals that word-level errorsprimarily originate from subtle phoneme inaccuracies amplifiedby downstream segmentation. Furthermore, a language modelweight sweep reveals a sharp phase transition: beyond a narrowtiebreaker role (wLM > 0.03), WER rises from 17% to 96%as the LM’s score scale overwhelms acoustic evidence. Thesefindings suggest that explicitly decoupling acoustic and lexicalprocessing yields interpretable error diagnostics and substantialparameter savings, at a moderate accuracy cost relative to endto-end models.