📄 Influence of Mother Tongue on English Accent

Summary

The paper asks whether a speaker's South Indian mother tongue (L1) can be detected from how they speak English (L2). The task is framed as a variant of Language Identification (LID) and speaker recognition. The authors recorded their own corpus of native speakers of Telugu, Tamil and Kannada, about one hour per language. They deliberately left out people from border regions so that dialect mixing would not blur the training data.

Each speech sample is converted into 13-dimensional MFCC vectors, and a Gaussian Mixture Model (GMM) is fitted for each language. Test material comes from a separate group of speakers who read English passages. Each test utterance gets its own model, and its GMM means are compared by distance with the three language models.

Pairwise confusion matrices show identification rates of about 80%, rising to 85% when acoustic-prosodic features are added. Tamil and Kannada were the hardest pair to separate, and prosody reduced that confusion. The work matters for accent-aware speaker verification and recognition in India's many-language setting, where such systems are still scarce.

Key points

• L1 languages are Telugu (tel), Tamil (tam) and Kannada (kan). The L2 is short English discourse. The models are GMMs over 13-dim MFCCs, plus acoustic-prosodic features in a second setup.
• Training corpus: 3600 s of native-language speech per language. Speakers: TEL 5M/4F, TAM 3M/3F, KAN 4M/4F. Recordings use general texts on personality development, read in relaxed conditions.
• Test set: English recordings of 30–90 s. Speakers: TEL 7M/7F, TAM 7M/5F, KAN 4M/8F. All are graduates, and English teachers and convent-educated speakers were excluded.
• Classification works by measuring the distance between the GMM means of a test utterance and each language model. Results are reported as pairwise confusion matrices, with a best of 85% versus 80% for cepstral features alone.
• Future work proposed: SDC features, the i-vector approach, and studying how other languages influence Hindi.

Interesting ideas

• I1. Models are trained only on L1 speech but applied to L2 speech, betting that native habits leak into the foreign language and can be measured there.
• I2. Speakers from border areas were removed from training and highly trained English speakers from testing, so each class stays clean and the test reflects ordinary accented usage.
• I3. Results are given as pairwise confusion matrices rather than one accuracy number, which exposes exactly which classes blur together (Tamil vs Kannada).
• I4. The same split is run twice, once with spectral features and once with prosodic features added, isolating what the extra signal contributes.
• I5. Mixture components are treated as stand-ins for broad phonetic classes, which avoids the need for phoneme-labelled data.

For Mova

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

• I1 (tokens): a similar "interference profile" could flag English text likely written by Ukrainian speakers (e.g., dropped articles, calqued prepositions), so the tagger can explain odd tokens instead of mis-tagging them.
• I2 (learn): when building silver data, drop sources whose genre or dialect is mixed, and keep atypical sources as a separate control set rather than mixing them in.
• I3 (learn): report parser and frame errors as confusion tables between labels or UD dialects, so repair rules can target the specific pairs that get mixed up.
• I4 (learn): ablate one knowledge source at a time (e.g., VerbNet roles, WordNet ontology) on a fixed test set to show its explainable contribution.
• I5 (world): unlabelled clusters over usage contexts could be inspected and named as coarse classes, giving a cheap first draft of categories before linking them to the ontology.