📄 NameDat: A Database of English Proper Names Spoken by Native Norwegians

Summary

Recognising proper names is hard for ASR, and foreign names are harder still. A Norwegian speaker saying an English name may mix native and English sounds to different degrees, and standard letter-to-sound rules predict these pronunciations poorly. Before this work, the only Norwegian resource for non-native names was Onomastica, a lexicon with one nativised transcription per name and no audio. The authors built NameDat, a small speech corpus of English names read inside Norwegian carrier sentences by native Norwegians. Names were chosen for three properties: how familiar they are likely to be, how badly an automatic recogniser is expected to handle them, and how many English-only phonemes they contain. An expert phonetician corrected machine-generated alignments in a multi-tier Praat annotation. In a recognition test, manually verified transcriptions gave lower error than automatic ones. Acoustic experiments showed that training non-native phones on TIMIT helps little compared with mapping them to the closest Norwegian phones, especially once triphones are used. The resource supports both lexical and acoustic modelling of non-native pronunciation variation for the SVoG recogniser.

Key points

• Speakers and data: 33 Norwegian speakers aged 18–60, balanced by age, gender, education and self-assessed English proficiency. Each read 125 sentences containing two English names, drawn from five manuscripts totalling 1250 unique names, mostly place names plus some US/UK person names.
• Recording: Sennheiser headset, 48 kHz, 16-bit PCM. The recording chain had a signal-to-noise ratio of 51 dBA. Sessions took place in a soundproof lab or an office, and speakers were told to say names the way they naturally would.
• Annotation: Norwegian SAMPA was extended with 11 British English phonemes. The Praat files have five tiers (auto, phone, phone comment, word, utterance), and one expert annotator did the work. At the time of writing, 19 of the 33 speakers were annotated.
• Phone recogniser results: non-native phones mapped to the nearest Norwegian phone ("nativized") scored 39.2%, versus 39.9% for phones trained on TIMIT. Adding a TIMIT-trained "r" alone brought this to 38.1%.
• Word recogniser results: with triphones, the gap almost disappeared, at 13.8% versus 13.5%. Monophones scored 19.5% versus 20.5%.

Interesting ideas

• I1. The authors picked "difficult" items by measuring the Levenshtein distance between an automatic transcription and an expert one, so the corpus deliberately over-samples the system's failure cases.
• I2. The annotator never started from scratch: machine output sat in an "auto" tier, and corrections went into a parallel tier, so the machine-to-human diff is preserved.
• I3. The annotation separated reading slips (marked as mispronunciations) from errors caused by limited knowledge of English (left unmarked as normal variation). This keeps noise apart from genuine variation.
• I4. The phone inventory gained only the units the native set lacked, and those defaulted to their closest native equivalent. A dedicated model was trained only for the one unit ("r") whose errors were widely scattered.
• I5. How familiar a name is likely to be was estimated from news-corpus frequency and, for cities, population size.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Rank sentences by edit distance between a witness parser's tree and a gold or teacher tree, then spend annotation and training effort on the most divergent ones.
• I2 (learn): Store the automatic parse next to every corrected one, so the stored diffs can feed directly into self-induced repair rules.
• I3 (tokens): When tagging learner or noisy text, give typos a different tag from systematic non-native forms, so the tagger can learn the latter without learning the former.
• I4 (tokens): Map foreign names and loanwords onto the nearest English morphology or POS pattern by default, and add special handling only for the classes where error analysis shows confusion.
• I5 (world): Use Wikidata population or corpus frequency as a prior on whether a reader is expected to know an entity, which can guide coreference and what an answer needs to explain.