---
title: "Whispering in Amharic — Gete 2025"
description: "Whisper performs poorly on Amharic because the language makes up only a tiny part of its multilingual training data."
authors: ["Dawit Ketema Gete", "Bedru Yimam Ahmed", "Tadesse Destaw Belay", "Yohannes Ayana Ejigu", "Sukairaj Hafiz Imam", "Alemu Belay Tessema", "Mohammed Oumer Adem", "Tadesse Amare Belay", "Robert Geislinger", "Umma Aliyu Musa", "Martin Semmann", "Shamsuddeen Hassan Muhammad", "Henning Schreiber", "Seid Muhie Yimam"]
year: 2025
venue: "arXiv:2503.18485"
source: "https://arxiv.org/abs/2503.18485"
doi: "10.48550/arxiv.2503.18485"
arxiv: "2503.18485"
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/gete-2025-whispering-amharic-fine-tuning-whisper/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/gete-2025-whispering-amharic-fine-tuning-whisper/)

[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/3/) / Whispering in Amharic: Fine-tuning Whisper for Low-resource Language

# Whispering in Amharic: Fine-tuning Whisper for Low-resource Language

**Authors:** Dawit Ketema Gete, Bedru Yimam Ahmed, [Tadesse Destaw Belay](https://md.research.elamur.ai/free/en/authors/tadesse-destaw-belay/), Yohannes Ayana Ejigu, [Sukairaj Hafiz Imam](https://md.research.elamur.ai/free/en/authors/sukairaj-hafiz-imam/), Alemu Belay Tessema, Mohammed Oumer Adem, Tadesse Amare Belay, [Robert Geislinger](https://md.research.elamur.ai/free/en/authors/robert-geislinger/), Umma Aliyu Musa, [Martin Semmann](https://md.research.elamur.ai/free/en/authors/martin-semmann/), [Shamsuddeen Hassan Muhammad](https://md.research.elamur.ai/free/en/authors/shamsuddeen-hassan-muhammad/) et al. (14 authors) · **Year:** 2025 · **Venue:** arXiv:2503.18485 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2503.18485) · [PDF](https://arxiv.org/pdf/2503.18485) · [DOI 10.48550/arxiv.2503.18485](https://doi.org/10.48550/arxiv.2503.18485) · [arXiv 2503.18485](https://arxiv.org/abs/2503.18485) · [OpenAlex](https://openalex.org/W6891879036)

**Citations:** 0 citations (OpenAlex, 2026-10-14) · [2](https://md.research.elamur.ai/free/en/papers/gete-2025-whispering-amharic-fine-tuning-whisper/cited-by/) in our library, of the 140884 papers whose bibliographies we read

## Summary
Whisper performs poorly on Amharic because the language makes up only a tiny part of its multilingual training data. The authors set out to fix this by adapting the Whisper-small checkpoint. They fine-tuned it on three public Amharic speech corpora: the Amharic part of Mozilla Common Voice v17.0, the Amharic portion of Google FLEURS, and the noisy BDU-speech corpus from Bahir Dar University. Several variants were trained on different combinations of these sources. Each was scored with WER, CER and BLEU, and the authors also inspected outputs by hand through a Gradio demo.

The strongest model, Whisper-small-am, came from training on FLEURS mixed with the newer corpora. Fine-tuning only on unfamiliar data gave weak results, while adding FLEURS, which Whisper had already seen in pretraining, helped stabilise learning. The authors also found that normalising Amharic homophone characters before scoring noticeably improves WER and BLEU. This suggests that unnormalised metrics overstate the real error rate. The work matters because it shows that the mix of training data and the evaluation conventions both strongly shape ASR quality for low-resource languages.

## Key points
- The base model is Whisper-small; Whisper itself was pretrained on about 680K hours of weakly supervised multilingual web audio.
- Training and evaluation data came from three sources:
  - Common Voice v17.0 (Amharic subset)
  - FLEURS Amharic (roughly 12 h per language in the full benchmark)
  - BDU-speech: about 44 h 46 min, 50 speakers reading 400 sentences each, clips of 4–20 s with heavy noise and non-native accents
- The released variants are whisper-small-am, whisper-small-fc-am, whisper-small-am-fleurs, whisper-small-am-common-speech, whisper-small-am-v2 and whisper-small-am-on-aggregated. They were compared on separate test sets, including the Common Voice test split.
- Audio was resampled to 16 kHz and passed through the Whisper feature extractor. Homophone normalisation was used only at evaluation time, not during training.
- The authors list several limitations:
  - only the small model size was tried
  - there was no comparison with SeamlessM4T or USM
  - data augmentation and continual learning were not explored
  - post-correction with NER, spell checkers or grammar checkers is proposed but not built

## Interesting ideas
- Mixing data the model already knows (FLEURS) into the fine-tuning set alongside new corpora keeps adaptation from collapsing, whereas training on new data alone hurt quality. — For Mova (learn): when retraining the perceptron parser or gparse on silver data from witness parsers, always mix in a share of the original gold UD treebank so new material does not erode what already works.
- Folding equivalent spellings into one canonical form before computing the metric keeps purely orthographic differences from being counted as errors. — For Mova (tokens): normalise British/American spellings, quote styles and hyphenation variants before scoring the tokenizer and lemmatizer, so error counts reflect real mistakes.
- A dedicated noisy, accented test set serves as a separate robustness probe rather than being blended into one average score. — For Mova (tree): keep a distinct evaluation slice of messy, typo-heavy or non-native English, and report parser LAS on it separately from clean treebank scores.
- Automatic scores are paired with manual inspection of outputs, because WER and BLEU cannot tell whether a transcript is usable or meaningful. — For Mova (answer): alongside accuracy on school tests and fable QA, regularly review a sample of answers by hand to catch replies that score well but rest on wrong reasoning.
- The authors suggest a correction layer after the main model, using spell checkers, NER and grammar tools, to fix leftover character-level errors. — For Mova (back): after generating text from the meaning graph, run a lightweight checking pass (agreement, entity names, spelling) that flags and repairs surface errors in an explainable way.

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

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/)

Topics: [Low-resource languages](https://md.research.elamur.ai/free/en/topics/low-resource-languages/) · [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/)

## Cite

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

```bibtex
@misc{gete2025whispering,
  title = {{Whispering in Amharic: Fine-tuning Whisper for Low-resource Language}},
  author = {Gete, Dawit Ketema and Ahmed, Bedru Yimam and Belay, Tadesse Destaw and Ejigu, Yohannes Ayana and Imam, Sukairaj Hafiz and Tessema, Alemu Belay and Adem, Mohammed Oumer and Belay, Tadesse Amare and Geislinger, Robert and Musa, Umma Aliyu and Semmann, Martin and Muhammad, Shamsuddeen Hassan and Schreiber, Henning and Yimam, Seid Muhie},
  year = {2025},
  doi = {10.48550/arxiv.2503.18485},
  eprint = {2503.18485},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2503.18485}
}
```

- [← previous Universal Dependencies Treebank for Uzbek](https://md.research.elamur.ai/free/en/papers/akhundjanova-2025-universal-dependencies-treebank-uzbek/)
- [next → A Classifier of Word-Level Variants in Witnesses of Biblical Hebrew Manuscripts](https://md.research.elamur.ai/free/en/papers/stoupak-2025-classifier-word-level-variants-witnesses/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/3/)

## References

11 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. 2012. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal processing magazine, 29(6):82–97. [find](https://scholar.google.com/scholar?q=Geoffrey+Hinton%2C+Li+Deng%2C+Dong+Yu%2C+George+E+Dahl%2C+Abdel%2Drahman+Mohamed%2C+Navdeep+Jaitly%2C+Andrew+Senior%2C+Vincent+Vanhoucke%2C+Patrick+Nguyen%2C+Tara+N+Sainath%2C+et+al%2E+2012%2E+Deep+neural+networks+for+acoustic)
2. Samuael Adnew and Paul Pu Liang. 2024. Semantically corrected amharic automatic speech recognition. arXiv.org, abs/2404.13362. [↗](https://arxiv.org/abs/2404.13362) [our card](https://md.research.elamur.ai/free/en/papers/adnew-2024-semantically-corrected-amharic-automatic-speech/)
3. Daniel Jurafsky and James H. Martin. 2000. Speech & language processing. Pearson Education India. [find](https://scholar.google.com/scholar?q=Daniel+Jurafsky+and+James+H%2E+Martin%2E+2000%2E+Speech+%26+language+processing%2E+Pearson+Education+India%2E)
4. Hamza Kheddar, Mustapha Hemis, and Yassine Himeur. 2024. Automatic speech recognition using advanced deep learning approaches: A survey. Information Fusion, page 102422. [find](https://scholar.google.com/scholar?q=Hamza+Kheddar%2C+Mustapha+Hemis%2C+and+Yassine+Himeur%2E+2024%2E+Automatic+speech+recognition+using+advanced+deep+learning+approaches%3A+A+survey%2E+Information+Fusion%2C+page+102422%2E)
5. Tesfa Tegegne Assfaw, Tsegaye Abebe, Belisty Yalew, and Tadesse Destaw Belay. 2022. Dialect-based noisy speech dataset, pre-processing tools, and recognition models for amharic. In 2022 International Conference on Information and Communication Technology for Development for Africa (ICT4DA), pages 90–95. IEEE. [find](https://scholar.google.com/scholar?q=Tesfa+Tegegne+Assfaw%2C+Tsegaye+Abebe%2C+Belisty+Yalew%2C+and+Tadesse+Destaw+Belay%2E+2022%2E+Dialect%2Dbased+noisy+speech+dataset%2C+pre%2Dprocessing+tools%2C+and+recognition+models+for+amharic%2E+In+2022+International+)
6. Yogesh Kumar. 2024. A comprehensive analysis of speech recognition systems in healthcare: Current research challenges and future prospects. SN Computer Science, 5(1):137. [find](https://scholar.google.com/scholar?q=Yogesh+Kumar%2E+2024%2E+A+comprehensive+analysis+of+speech+recognition+systems+in+healthcare%3A+Current+research+challenges+and+future+prospects%2E+SN+Computer+Science%2C+5%281%29%3A137%2E)
7. Tadesse Destaw Belay, Abinew Ali Ayele, Getie Gelaye, Seid Muhie Yimam, and Chris Biemann. 2021. Impacts of homophone normalization on semantic models for amharic. In 2021 International Conference on Information and Communication Technology for Development for Africa (ICT4DA), pages 101–106. IEEE. [find](https://scholar.google.com/scholar?q=Tadesse+Destaw+Belay%2C+Abinew+Ali+Ayele%2C+Getie+Gelaye%2C+Seid+Muhie+Yimam%2C+and+Chris+Biemann%2E+2021%2E+Impacts+of+homophone+normalization+on+semantic+models+for+amharic%2E+In+2021+International+Conference+on+)
8. Per E Kummervold, Javier de la Rosa, Freddy Wetjen, Rolv-Arild Braaten, and Per Erik Solberg. 2024. Whispering in norwegian: Navigating orthographic and dialectic challenges. arXiv preprint arXiv:2402.01917. [↗](https://arxiv.org/abs/2402.01917)
9. Jinpeng Li, Yu Pu, Qi Sun, and Wei-Qiang Zhang. 2024a. Improving whisper’s recognition performance for under-represented language kazakh leveraging unpaired speech and text. arXiv preprint arXiv:2408.05554. [↗](https://arxiv.org/abs/2408.05554)
10. Tianyang Zhong, Zhenyuan Yang, Zheng Liu, Ruidong Zhang, Yiheng Liu, Haiyang Sun, Yi Pan, Yiwei Li, Yifan Zhou, Hanqi Jiang, Junhao Chen, and Tianming Liu. 2024. Opportunities and challenges of large language models for low-resource languages in humanities research. ArXiv, abs/2412.04497. [↗](https://arxiv.org/abs/2412.04497)
11. Jinpeng Li, Yu Pu, Qi Sun, and Wei-Qiang Zhang. 2024b. Improving whisper’s recognition performance for under-represented language kazakh leveraging unpaired speech and text. [find](https://scholar.google.com/scholar?q=Jinpeng+Li%2C+Yu+Pu%2C+Qi+Sun%2C+and+Wei%2DQiang+Zhang%2E+2024b%2E+Improving+whisper%E2%80%99s+recognition+performance+for+under%2Drepresented+language+kazakh+leveraging+unpaired+speech+and+text%2E)