DIALECT-AWARE SPEECH RECOGNITION AND GRAMMATICAL NORMALIZATION FOR TURKIC LANGUAGES: A HYBRID MODEL AND ALGORITHMS
Keywords:
Keywords: automatic speech recognition; dialect; phonetic variability; grammatical normalization; Wav2Vec2 XLS-R; KenLM; Turkic languages; Uzbek; low-resource languages.Abstract
Abstract: Automatic speech recognition (ASR) systems are trained mostly on large, relatively homogeneous speech corpora, so their accuracy drops on speech varieties that are under-represented in the training data. This problem is especially acute for Turkic languages, where dialects typically have no standard orthography and no annotated corpora. We review international evidence on language variation, language endangerment and unequal ASR accuracy across speech varieties, and propose a hybrid model for dialect-aware recognition and grammatical normalization. The model has three components: (1) a multilingual pretrained Wav2Vec2 XLS-R acoustic model fine-tuned with CTC; (2) weighted phonetic correspondence rules that map dialectal pronunciation to the standard form; and (3) a normalization stage that combines KenLM n-gram rescoring with grammatical error detection. We give a formal description, the algorithm and an evaluation protocol based on word and character error rates (WER, CER) measured separately for standard and dialectal speech. This is a methodological paper: experimental results will be reported in follow-up work.
References
[1] Merriam-Webster. Merriam-Webster's Collegiate Dictionary, 12th ed. September 2025. https://www.merriam-webster.com
[2] Oxford English Dictionary. January 2023 update: Indian English pronunciations. https://www.oed.com
[3] Koenecke, A., Nam, A., Lake, E., et al. Racial disparities in automated speech recognition. Proceedings of the National Academy of Sciences, 117(14), 7684–7689 (2020).
[4] Vaani: Capturing the Language Landscape for an Inclusive Digital India. IISc and ARTPARK. https://vaani.iisc.ac.in
[5] Ethnologue. How many languages are there in the world? https://ethnologue.com/insights/how-many-languages
[6] UNESCO. International Decade of Indigenous Languages 2022–2032. https://www.unesco.at/en/cross-sectional-topics/article/unesco-celebrates-the-international-decade-of-indigenous-languages
[7] Mukhamadiyev, A., Khujayarov, I., Djuraev, O., Cho, J. Automatic speech recognition method based on deep learning approaches for Uzbek language. Sensors, 22(10), 3683 (2022).
[8] Babu, A., Wang, C., Tjandra, A., et al. XLS-R: Self-supervised cross-lingual speech representation learning at scale. arXiv:2111.09296 (2021).
[9] Ardila, R., Branson, M., Davis, K., et al. Common Voice: A massively-multilingual speech corpus. Proceedings of LREC, 4218–4222 (2020).
[10] Musaev, M., Mussakhojayeva, S., Khujayorov, I., Khassanov, Y., Ochilov, M., Varol, H.A. USC: An open-source Uzbek speech corpus and initial speech recognition experiments. SPECOM (2021).
[11] Graves, A., Fernández, S., Gomez, F., Schmidhuber, J. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. Proceedings of ICML, 369–376 (2006).
[12] Heafield, K. KenLM: Faster and smaller language model queries. Proceedings of the Sixth Workshop on Statistical Machine Translation, 187–197 (2011).
