--- title: "Dialect-Specific Models for Automatic Speech Recognition of" description: "This student paper tests whether training on dialect-matched data helps African American Vernacular… 1 reference, cited by 10 papers here." authors: ["Rachel Dorn"] year: 2019 venue: "Proceedings of the Student Research Workshop Associated with RANLP 2019" source: "https://aclanthology.org/R19-2003/" doi: "10.26615/issn.2603-2821.2019_003" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/dorn-2019-dialect-specific-models-automatic-speech/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/dorn-2019-dialect-specific-models-automatic-speech/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2019](https://md.research.elamur.ai/free/en/papers/2019/page/6/) / Dialect-Specific Models for Automatic Speech Recognition of African American Vernacular English # Dialect-Specific Models for Automatic Speech Recognition of African American Vernacular English **Authors:** Rachel Dorn · **Year:** 2019 · **Venue:** [Proceedings of the Student Research Workshop Associated with RANLP 2019](https://md.research.elamur.ai/free/en/venues/ranlp/) ([RANLP 2019](https://md.research.elamur.ai/free/en/venues/ranlp/2019/)) · **Publisher:** [INCOMA Ltd.](https://md.research.elamur.ai/free/en/publishers/incoma-ltd/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/R19-2003/) · [PDF](https://aclanthology.org/R19-2003.pdf) · [DOI 10.26615/issn.2603-2821.2019\_003](https://doi.org/10.26615/issn.2603-2821.2019_003) · [OpenAlex](https://openalex.org/W2982448527) **Citations:** 20 citations (OpenAlex, 2026-10-14) · [10](https://md.research.elamur.ai/free/en/papers/dorn-2019-dialect-specific-models-automatic-speech/cited-by/) in our library, of the 149,439 papers whose bibliographies we read ## Summary Speech recognizers tend to perform worse for speakers of non-standard dialects, partly because widely used corpora contain very few such voices. For example, only 26 of TIMIT's 630 speakers are black, and Switchboard lists no AAVE dialect at all. This student paper tests whether training on dialect-matched data helps African American Vernacular English (AAVE) speakers. Using interview recordings from CORAAL, the author built three models with Mozilla's DeepSpeech: - an AAVE model trained on the interviewees, - an SAE model trained on the interviewers, - a combined baseline trained on everything. Each test set was transcribed by its matching model, by the opposite-dialect model, and by the baseline. Output was scored with word error rate (WER) and with character-level Levenshtein distance. Absolute accuracy was low; WER exceeded 1 because of spurious word breaks. Even so, matched-dialect models beat the pooled baseline for both dialects. The AAVE model improved by more than 18% on Levenshtein distance. Because this setup isolates the effect of including AAVE speakers in training data, it gives concrete evidence that corpus diversity matters for fair ASR. ## Key points - **Data:** a Washington D.C. subset of CORAAL (University of Oregon) was used to limit regional variation within the dialect. It contains 42 adult speakers (6 interviewers, 36 interviewees aged 18–77) and 31,468 utterances, averaging about four seconds each. - **Splits:** a 70-20-10 train/dev/test split kept each speaker proportionally represented. - AAVE: 15,857 / 4,467 / 2,266 utterances. - SAE: 6,199 / 1,783 / 902 utterances. - Average utterance length was 6.2 words (AAVE) versus 4.6 words (SAE). - **Preprocessing:** audio was converted to 16 kHz mono WAV. Transcripts were lowercased, stripped of punctuation, had numerals spelled out, and had non-verbal markers removed with regex. - **Results on AAVE test utterances:** - WER: 0.9363 (AAVE model), 1.1509 (combined), 1.1886 (SAE model). - Levenshtein distance: 15.28, 18.75 and 21.16 respectively. - **Results on SAE test utterances and relative gains:** - Levenshtein distance: 13.00 (SAE model), 15.06 (combined), 15.55 (AAVE model). - WER improvement over the baseline was 16.6% for AAVE and 7.5% for SAE. Many SAE utterances were short backchannels such as "mm hm". ## Interesting ideas - Pairing word-level WER with a character-level edit distance gives partial credit when a model gets sounds or fragments right inside a wrongly recognized word, revealing progress that WER hides. - Running each specialised model on the other group's test data separates "this data is harder" from "this model fits its own data better". - Splitting train/dev/test so every speaker appears proportionally in each part prevents one dominant source from skewing results. - A small dialect-matched model outperforming a larger pooled model suggests that mixing varieties can blur what a model learns. - The paper's table of AAVE phonetic features (dropped final consonants, "-in'" endings, "da" for "the") lists regular, predictable spelling variants of standard forms. ## For Mova How these ideas could serve Mova, our small open language model — its files, stages and tests: - **I1** (tokens): score lemmatization and morphology on noisy or non-standard input with a character-level distance alongside exact-match accuracy, so near-misses on eye-dialect forms are visible. - **I2** (learn): adopt cross-domain swaps as a standard negative control, e.g. a parser trained on one register evaluated on another, before claiming a domain-specific gain. - **I3** (learn): stratify silver and gold splits by document or author so evaluation is not driven by a few heavy contributors. - **I4** (tree): consider dialect-tagged training subsets or a dialect-aware UD conversion layer for gparse instead of pooling all English text, and check whether per-variety accuracy improves. - **I5** (tokens): encode such systematic variants as explainable normalization rules in the tokenizer/lemmatizer, mapping "walkin'" to the lemma "walk" while keeping the surface form for downstream use. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* ## Institutions Also on the byline: [Virginia Commonwealth University](https://ror.org/02nkdxk79). 📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works 🧭 Topics (2): [English](https://md.research.elamur.ai/free/en/topics/english/) · 2,263 works · [🎙️ Speech & spoken language](https://md.research.elamur.ai/free/en/topics/speech/) · 480 works ## Cite Cite as: Rachel Dorn. 2019. Dialect-Specific Models for Automatic Speech Recognition of African American Vernacular English. Proceedings of the Student Research Workshop Associated with RANLP 2019. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{dorn2019dialect, title = {{Dialect-Specific Models for Automatic Speech Recognition of African American Vernacular English}}, author = {Rachel Dorn}, booktitle = {Proceedings of the Student Research Workshop Associated with RANLP 2019}, year = {2019}, doi = {10.26615/issn.2603-2821.2019\_003}, url = {https://doi.org/10.26615/issn.2603-2821.2019_003} } ``` - [← previous Topic Tensor Network for Implicit Discourse Relation Recognition in Chinese](https://md.research.elamur.ai/free/en/papers/xu-2019-topic-tensor-network-implicit-discourse/) - [next → Do You Know That Florence Is Packed with Visitors? Evaluating State-of-the-art Models of Speaker Commitment](https://md.research.elamur.ai/free/en/papers/jiang-2019-do-you-know-that-florence/) - [↑ back to papers of 2019](https://md.research.elamur.ai/free/en/papers/2019/page/6/) ## References 1 references, 0 of them with a card of ours. ↗ leads to the work. 1. Horia Cucu, Laurent Besacier, Corneliu Burileanu, and Andi Buzo. 2012. Asr domain adaptation methods for low-resourced languages: Application to romanian language. In 20th European Signal Processing Conference. pages 1648–1652. Levenshtein Distance 18.5% 13.7% Mohamed Elmahdy, Mark Hasegawa-Johnson, and Eiman Mustafawi. 2013. A transfer learning approach for under-resourced arabic dialects speech recognition. In Workshop on Less Resourced Languages, new technologies, new challenges and opportunities (LTC 2013). pages 60–64. *Table 3: Improvements in Error Rate Between Dialect-Specific Model and Combined Model* John S. Garofolo, Lori F. Lamel, William M. Fisher, Jonathan G. Fiscus, David S. Pallett, Nancy L. Dahlgren, and Victor Zue. 1993. Timit acoustic-phonetic continuous speech corpus ldc93s1. Linguistic Data Consortium https://catalog.ldc.upenn.edu/docs/LDC93S1. Conclusions and Future Work [↗](https://doi.org/10.5281/zenodo.52416) *Updated 1 November 2026*