Selective Augmentation: Improving Universal Automatic Phonetic Transcription via G2P Bootstrapping

Selective Augmentation introduces a novel bootstrapping technique for universal automatic phonetic transcription (APT) that leverages grapheme-to-phoneme (G2P) mappings to enrich training data across languages. The key insight: instead of manually curating massive datasets with fine-grained phonetic distinctions, researchers can “borrow” distinctions from a helper language—in this case, Hindi—to augment transcriptions in a target language (German). Applied to the MultIPA model, the method corrected voicing errors: false positives dropped, yielding a 17.6% boost in voicing accuracy. More strikingly, it introduced aspiration recognition where none existed—baseline model transcribed 0% of German /p, t, k/ as aspirated, but Selective Augmentation raised that to 61.2%. This also reduced conflations between tenuis and aspirated plosives by 32.2%, demonstrating that cross-lingual transfer can systematically improve phonetic fidelity.

The approach addresses a fundamental bottleneck in speech technology: high-quality phonetic transcriptions are expensive and scarce. By automating the transfer of phonetic distinctions from languages that mark them explicitly (like Hindi's aspirated/unaspirated contrast), the method scales to under-resourced languages. The paper, accepted at LREC 2026, includes objective metrics to validate success, overcoming the intrinsic challenges of evaluating phonetic transcription quality. For practitioners, this means more accurate speech recognition, language learning tools, and dialect analysis without manual labeling. Selective Augmentation offers a practical path to universal phonetic transcription that adapts to new languages and features with minimal human effort.