📄 Experimental detection of vowel pronunciation variants in Amharic

Summary

Building a pronunciation lexicon for speech recognition usually depends on expert knowledge of the language. For Amharic such expertise is scarce. Pellegrini and Lamel (LIMSI-CNRS) use the recognizer itself to discover which vowel variants actually occur. Their data is 37 hours of radio broadcasts transcribed by native speakers, about 240k words with 50k distinct lexemes.

Because this corpus is too small for reliable word-level statistics, they first work with syllables. Each syllable in the lexicon may keep its vowel, drop it, or swap it for any other vowel. Forced alignment then shows which realizations the acoustic models prefer. The most frequent alternate for each syllable onset is turned into word-level variants, which are checked with context-independent and context-dependent models using the variant2+ rate.

The main findings are that schwas are mostly not pronounced and that vowels pair up into confusion classes. Words containing a schwa are aligned with reduced forms far more often than other words. The approach matters for any language with little data and limited linguistic expertise.

Key points

• The 240 Amharic script symbols are transliterated to 33 phoneme characters, which also merges homophone symbols. The phone set is 33 phones plus silence, breath and hesitation units.
• In syllable-level alignment, schwa (x) is deleted about 61% of the time. /E/ keeps its identity in under half of its occurrences.
• The dominant vowel confusions are /E/–/a/, /e/–/i/ and /o/–/u/. The authors relate these to the three-vowel system of early Semitic languages and Arabic.
• The word lexicon allows one variant per syllable, which gives 2^N forms for an N-syllable word. With 100 CD models, the most frequent word nEwx has a variant2+ rate of 90.3% and mEto has 49.5%. Words with schwas average 86.9% variant2+ versus 51.3% for words without.
• The acoustic model sets were 36 CI models and 100/300/1k/7.2k triphone CD models. The CI models gave the highest variant2+ and the 7.2k CD models the lowest. Variant2+ grows with word frequency rank, along with word length and the share of words containing a schwa.

Interesting ideas

• I1. Deliberately overgenerate alternatives (every vowel substitution or deletion), let a trained model choose among them in forced alignment, and keep only the alternates it frequently selects; this turns a model into a data-driven linguistic probe.
• I2. When word-level counts are too sparse, gather statistics on a smaller recurring unit and recompose; boundary markers on each unit keep recombination into words unambiguous.
• I3. Cap the number of alternates per unit at one, so short frequent items still get variants while long items stay tractable (2^N rather than a combinatorial blowup).
• I4. Measure the need for variants as the share of instances where the system picks something other than the canonical form (variant2+), and break it down by frequency band.
• I5. Two biases can hide variation: measuring on the training data, and using a richer model that absorbs variability internally (CD models lowered variant2+).

For Mova

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

• I1 (learn): candidate repair rules or alternate analyses could be overgenerated, with witness parsers voting among them, and only alternates chosen often across a corpus promoted to rules.
• I2 (tokens): morphology statistics (e.g. affix or spelling variants) could be estimated per subword unit with explicit boundary marks, then projected back to whole lemmas.
• I3 (tree): ambiguity packing in the parser or tagger could allow only the single most frequent alternate per token to keep the candidate space bounded and explainable.
• I4 (learn): a "non-default rate" per construction or frame, split by frequency, would show where the canonical analysis fails most and where to focus repair.
• I5 (learn): any rule-need statistic should be computed on held-out text and compared across model strengths, as a control against overly optimistic conclusions.