Phonologically Informed Tokenization for German Speech Recognition: A Cross-Domain Study

Key points

• German is a morphologically rich language whose syllable structure is exceptionally well-predicted by the Knuth--Liang hyphenation algorithm.1
• We compare three tokenizer families on the Omnilingual ASR wav2vec 2.0 backbone fine-tuned with CTC: the pretrained multilingual character inventory, a data-driven Byte-Pair Encoding (BPE) over orthography, and phonologically informed units from Pyphen syllabification and grapheme-to-phoneme conversion.1
• Across 40 fine-tunes, we evaluate on three German test sets spanning orthogonal shifts: in-domain read speech, dialectal spontaneous speech, and standard-German spontaneous speech.1
• Under domain shift the picture splits along vocabulary size rather than the linguistic axis of variation: at small vocabularies, syllable-aware tokenization improves on dialectal speech, where phonetic surface forms vary but syllable structure is preserved, and stays ahead on spontaneous speech, where new word-forms violate vocabulary closure.1

Sources (1)

• [1]Phonologically Informed Tokenization for German Speech Recognition: A Cross-Domain Study
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 8, 10:20 AM

  We ask whether phonologically informed tokenization can serve as a competitive target for end-to-end speech recognition.

  German is a morphologically rich language whose syllable structure is exceptionally well-predicted by the Knuth--Liang hyphenation algorithm.

Extractive summary: sentences quoted from the sources.