---
title: "Domain-Aware Speaker Diarization On African-Accented English"
description: "The paper asks whether speaker diarization (\"who spoke when\") performs worse on African-accented English in a clinical… 18 references."
authors: ["Chibuzor Okocha", "Kelechi Ezema", "Christan Grant"]
year: 2025
venue: "arXiv:2509.21554"
source: "https://arxiv.org/abs/2509.21554"
doi: "10.48550/arxiv.2509.21554"
arxiv: "2509.21554"
licence: "CC BY 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/okocha-2025-domain-aware-speaker-diarization-african/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/okocha-2025-domain-aware-speaker-diarization-african/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/16/) / Domain-Aware Speaker Diarization On African-Accented English

# Domain-Aware Speaker Diarization On African-Accented English

**Authors:** [Chibuzor Okocha](https://md.research.elamur.ai/free/en/authors/chibuzor-okocha/), Kelechi Ezema and [Christan Grant](https://md.research.elamur.ai/free/en/authors/christan-grant/) · **Year:** 2025 · **Venue:** arXiv:2509.21554 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2509.21554) · [PDF](https://arxiv.org/pdf/2509.21554) · [DOI 10.48550/arxiv.2509.21554](https://doi.org/10.48550/arxiv.2509.21554) · [arXiv 2509.21554](https://arxiv.org/abs/2509.21554) · [OpenAlex](https://openalex.org/W7083677820)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
The paper asks whether speaker diarization ("who spoke when") performs worse on African-accented English in a clinical setting than in everyday talk, and why. The authors take a subset of AfriSpeech-Dialog: two-speaker recordings of OSCE-style doctor–patient role plays and open-topic chats, some with code-switching. They run eight diarization systems on it, four commercial APIs and four open models, all in default mode. Scoring is deliberately unforgiving: one-to-one Hungarian speaker mapping, zero collar, and overlapping speech included.

Every system scores clearly worse on clinical audio, and the difference is statistically significant. Splitting the error into its parts shows the clinical penalty comes mostly from false alarms, then missed speech. This fits the short turns and overlap typical of clinical dialogue.

As a test of adaptation, they fine-tune only the pyannote segmentation module on AfriSpeech-Countries. This halves its error but leaves part of the domain gap. The authors conclude that the clinical difficulty is not just an accent or data-scarcity issue.

## Key points
- Eight systems were compared: AssemblyAI, Deepgram, Soniox, Rev.ai/Reverb (commercial) and pyannote-audio, CAM++, Sortformer, Titanet-L (open).
- The evaluation data has 20 medical conversations (2.07 h) and 29 general ones (4.93 h). AssemblyAI had the lowest overall DER (12.72%); Titanet-L was the best open model (16.27%).
- Mean DER was 33.38% on Medical vs 15.18% on General. The gap held under both a paired t-test (t(7) = 10.67, p \< 10⁻⁴) and a Wilcoxon test (W = 0, p = 0.0078).
- The fine-tuning recipe:
  - Target: pyannote/segmentation-3.0, with speaker embeddings frozen.
  - Training data: about 67.7 h of AfriSpeech-Countries clips, dominated by Nigeria, Kenya and South Africa.
  - Setup: frame-level BCE, 10 s chunks, Adam at 1e-4, 10 epochs, no augmentation.
  - Result: pyannote DER fell from 21.30% to 10.65%, and its Medical–General gap shrank from 12.43 to 6.21 p.p.
- Error components from General to Medical: false alarms rose most (+11.0 p.p.), misses less (about +5.9 p.p.), confusion only slightly (+1.9 p.p.). False alarms make up 54% of clinical DER.

## Interesting ideas
- Measure the domain gap per system on paired data and back it with two significance tests (parametric plus rank-based), rather than eyeballing averages.
- Break one headline error number into components (spurious output, missed material, wrong assignment) and report each one's share, because the mix shows which mechanism is failing.
- Adapt only one module while freezing the rest, then check whether the gap survives; a remaining gap suggests the difficulty is intrinsic to the domain, not missing training data.
- Use a strict scoring protocol that keeps the hard cases (zero tolerance, overlaps counted), since lenient scoring hides exactly where systems break.
- Profile each item structurally (turn count, utterance length, overlap) and relate those profiles to error, to explain failures by input shape rather than by label alone.

## For Mova

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

- **I1** (learn): Score each witness parser on genre splits of the same treebank, e.g. dialogue vs narrative vs fables, and test whether the gap is systematic before blaming a particular model.
- **I2** (tree): Split LAS losses into tokenization-induced errors, wrong head and wrong label, so repair efforts go to the component that actually grows on hard text.
- **I3** (learn): After retraining gparse on in-genre silver data, compare the leftover genre gap to separate data scarcity from structural difficulty such as fragments or ellipsis.
- **I4** (answer): Evaluate the readers without excluding multi-clause or ambiguous questions, so reported accuracy reflects the real hard tail.
- **I5** (frames): Record per-text features like turn-taking density or sentence fragmentation and correlate them with frame and coreference errors to find the conditions where pragmatics breaks.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

📚 Shelves (5): [🖥️ 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 · [🔊 Signal Processing](https://md.research.elamur.ai/free/en/shelves/computer-science/signal-processing/) · 452 works

🧭 Topics (1): [English](https://md.research.elamur.ai/free/en/topics/english/) · 2,263 works

## Cite

Cite as: Chibuzor Okocha, Kelechi Ezema and Christan Grant. 2025. Domain-Aware Speaker Diarization On African-Accented English. arXiv:2509.21554.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@misc{okocha2025domain,
  title = {{Domain-Aware Speaker Diarization On African-Accented English}},
  author = {Chibuzor Okocha, Kelechi Ezema and Christan Grant},
  year = {2025},
  doi = {10.48550/arxiv.2509.21554},
  eprint = {2509.21554},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2509.21554}
}
```

- [← previous Domain-Adaptive Pre-Training for Arabic Aspect-Based Sentiment Analysis: A Comparative Study of Domain Adaptation and Fine-Tuning Strategies](https://md.research.elamur.ai/free/en/papers/alyami-2025-domain-adaptive-pre-training-arabic/)
- [next → Editorial: Transactional and transnational contact within World Englishes: the Gulf region and beyond](https://md.research.elamur.ai/free/en/papers/lorenz-2026-editorial-transactional-transnational-contact-within/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/16/)

## References

18 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. [1] Tae Jin Park, Naoyuki Kanda, Dimitrios Dimitriadis, Kyu J Han, Shinji Watanabe, and Shrikanth Narayanan, “A review of speaker diarization: Recent advances with deep learning,” Computer Speech & Language, vol. 72, pp. 101317, 2022. [↗](https://doi.org/10.1016/j.csl.2021.101317)
2. [2] Aonan Zhang, Quan Wang, Zhenyao Zhu, John Paisley, and Chong Wang, “Fully supervised speaker diarization,” in ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). [find](https://scholar.google.com/scholar?q=%5B2%5D+Aonan+Zhang%2C+Quan+Wang%2C+Zhenyao+Zhu%2C+John+Paisley%2C+and+Chong+Wang%2C+%E2%80%9CFully+supervised+speaker+diarization%2C%E2%80%9D+in+ICASSP+2019%2D2019+IEEE+International+Conference+on+Acoustics%2C+Speech+and+Signal+Process)
3. [3] Tobi Olatunji, Busayo Awobade, Mardhiyah Sanni, Tassallah Abdullahi, Chibuzor Okocha, Kelechi Ezema, Devendra Deepak Kayande, Lukman Enegi Ismaila, and Gloria Ashiya Katuka, “Afrivox: Probing multilingual and accent robustness of speech llms,” 2025, Downloadable PDF available. [find](https://scholar.google.com/scholar?q=%5B3%5D+Tobi+Olatunji%2C+Busayo+Awobade%2C+Mardhiyah+Sanni%2C+Tassallah+Abdullahi%2C+Chibuzor+Okocha%2C+Kelechi+Ezema%2C+Devendra+Deepak+Kayande%2C+Lukman+Enegi+Ismaila%2C+and+Gloria+Ashiya+Katuka%2C+%E2%80%9CAfrivox%3A+Probing+mult)
4. [10] Soniox, “Soniox-7b-v1.0,” [https://huggingface.co/soniox/Soniox-7B-v1.0](https://huggingface.co/soniox/Soniox-7B-v1.0), 2025, Accessed: 2025-05-28. [↗](https://huggingface.co/soniox/Soniox-7B-v1.0)
5. [11] Nishchal Bhandari, Danny Chen, Miguel Ángel del Rı́o Fernández, Natalie Delworth, Jennifer Drexler Fox, Migüel Jetté, Quinten McNamara, Corey Miller, Ondřej Novotný, Ján Profant, Nan Qin, Martin Ratajczak, and Jean-Philippe Robichaud, “Reverb: Open-source asr and diarization from rev,” 2024. [↗](https://doi.org/10.48550/arxiv.2410.03930)
6. [12] Hervé Bredin, “pyannote. audio 2.1 speaker diarization pipeline: principle, benchmark, and recipe,” in 24th INTERSPEECH Conference (INTERSPEECH 2023). [find](https://scholar.google.com/scholar?q=%5B12%5D+Herve%CC%81+Bredin%2C+%E2%80%9Cpyannote%2E+audio+2%2E1+speaker+diarization+pipeline%3A+principle%2C+benchmark%2C+and+recipe%2C%E2%80%9D+in+24th+INTERSPEECH+Conference+%28INTERSPEECH+2023%29%2E)
7. [13] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur, “Librispeech: an asr corpus based on public domain audio books,” in 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). [find](https://scholar.google.com/scholar?q=%5B13%5D+Vassil+Panayotov%2C+Guoguo+Chen%2C+Daniel+Povey%2C+and+Sanjeev+Khudanpur%2C+%E2%80%9CLibrispeech%3A+an+asr+corpus+based+on+public+domain+audio+books%2C%E2%80%9D+in+2015+IEEE+international+conference+on+acoustics%2C+speech+and)
8. [4] Mardhiyah Sanni, Tassallah Abdullahi, Devendra D Kayande, Emmanuel Ayodele, Naome A Etori, Michael S Mollel, Moshood Yekini, Chibuzor Okocha, Lukman E Ismaila, Folafunmi Omofoye, et al., “[Afrispeech-dialog: A benchmark dataset for spontaneous english conversations in healthcare and beyond](https://md.research.elamur.ai/free/en/papers/sanni-2025-afrispeech-dialog-benchmark-dataset-spontaneous/),” arXiv preprint arXiv:2502.03945, 2025. [↗](https://arxiv.org/abs/2502.03945)
9. [14] Hui Wang, Siqi Zheng, Yafeng Chen, Luyao Cheng, and Qian Chen, “Cam++: A fast and efficient network for speaker verification using context-aware masking,” arXiv preprint arXiv:2303.00332, 2023. [↗](https://arxiv.org/abs/2303.00332)
10. [5] Tom McArthur, “World english and world englishes: Trends, tensions, varieties, and standards,” Language teaching, vol. 34, no. 1, pp. 1–20, 2001. [↗](https://doi.org/10.1017/s0261444800016062)
11. [16] Neville Ryant, Prachi Singh, Venkat Krishnamohan, Rajat Varma, Kenneth Church, Christopher Cieri, Jun Du, Sriram Ganapathy, and Mark Liberman, “The third dihard diarization challenge,” in Proc. Interspeech, 2021, arXiv:2012.01477. [↗](https://arxiv.org/abs/2012.01477)
12. [6] Tobi Olatunji, Tejumade Afonja, Aditya Yadavalli, Chris Chinenye Emezue, Sahib Singh, Bonaventure FP Dossou, Joanne Osuchukwu, Salomey Osei, Atnafu Lambebo Tonja, Naome Etori, et al., “[Afrispeech-200: Pan-african accented speech dataset for clinical and general domain asr](https://md.research.elamur.ai/free/en/papers/olatunji-2023-afrispeech-200-pan-african-accented/),” Transactions of the Association for Computational Linguistics, vol. 11, pp. 1669– 1685, 2023. [↗](https://doi.org/10.1162/tacl_a_00627)
13. [7] Zili Huang, Shinji Watanabe, Yusuke Fujita, Paola Garcı́a, Yiwen Shao, Daniel Povey, and Sanjeev Khudanpur, “Speaker diarization with region proposal network,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). [↗](https://doi.org/10.48550/arxiv.2002.06220)
14. [8] AssemblyAI, “Assemblyai: Ai models to transcribe and understand speech,” [https://www.assemblyai.com/](https://www.assemblyai.com/), 2025, Accessed: 2025-05-28. [↗](https://www.assemblyai.com/)
15. [9] Deepgram, “Model options,” [https://developers.deepgram.com/docs/model](https://developers.deepgram.com/docs/model), 2025, Accessed: 2025-05-28. [↗](https://developers.deepgram.com/docs/model)
16. [15] Claude Barras, Xuan Zhu, Sylvain Meignier, and Jean-Luc Gauvain, “Improving speaker diarization,” 2004. [find](https://scholar.google.com/scholar?q=%5B15%5D+Claude+Barras%2C+Xuan+Zhu%2C+Sylvain+Meignier%2C+and+Jean%2DLuc+Gauvain%2C+%E2%80%9CImproving+speaker+diarization%2C%E2%80%9D+2004%2E)
17. [17] Yusuke Fujita, Naoyuki Kanda, Shota Horiguchi, Kenji Nagamatsu, and Shinji Watanabe, “End-to-end neural speaker diarization with permutation-free objectives,” in Proc. Interspeech, 2019. [↗](https://doi.org/10.48550/arxiv.1909.05952)
18. [18] David Snyder, Daniel Garcia-Romero, Gregory Sell, Daniel Povey, and Sanjeev Khudanpur, “X-vectors: Robust dnn embeddings for speaker recognition,” in ICASSP, 2018. [find](https://scholar.google.com/scholar?q=%5B18%5D+David+Snyder%2C+Daniel+Garcia%2DRomero%2C+Gregory+Sell%2C+Daniel+Povey%2C+and+Sanjeev+Khudanpur%2C+%E2%80%9CX%2Dvectors%3A+Robust+dnn+embeddings+for+speaker+recognition%2C%E2%80%9D+in+ICASSP%2C+2018%2E)

*Updated 1 November 2026*