This model is part of WAXALNet, a suite of ASR models fine-tuned on the WAXAL corpus across 19 African languages, developed as part of the WAXAL ASR Benchmark study.

Model Details


Language	Amharic (`amh`)
Language Family	Afro-Asiatic
Architecture	MMS-300M (300M parameters)
Base Model	facebook/mms-300m
Training Data	WAXAL corpus (conversational spontaneous speech)
Test WER	37.8%
Test CER	13.2%
License	cc-by-nc-4.0

Intended Use

This model is intended for automatic speech recognition of Amharic conversational speech. It was evaluated on the WAXAL test set (spontaneous, image-prompted speech) and partially on FLEURS (read speech). It is suitable for research and low-resource ASR applications. It is not recommended for high-stakes production use without further validation.

Training Data

Fine-tuned on the WAXAL corpus, a large-scale dataset of transcribed, image-prompted spontaneous speech across 19 African languages recorded in participants' natural environments. The Amharic training split contains conversational speech across diverse speakers. Data is released under CC-BY 4.0.

Usage

python

from transformers import pipeline

asr = pipeline("automatic-speech-recognition",
               model="waxal-benchmarking/mms-300m-waxal-amh")
result = asr("audio.wav")
print(result["text"])

Test Set Performance (WAXAL Benchmark)

Evaluated on the filtered WAXAL test set (duration >= 1.5s, speech rate >= 4 WPS).

Metric	Score
WER	37.8%
CER	13.2%

Full benchmark results across all 19 languages and 6 models are reported in the WAXAL ASR Benchmark paper (citation below).

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0001
train_batch_size: 16
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 2
total_train_batch_size: 32
optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 500
num_epochs: 30
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
8.2224	0.4209	500	4.1240	1.0	1.0
8.1967	0.8418	1000	4.0965	1.0	1.0
8.1930	1.2626	1500	4.0816	1.0	1.0
8.0075	1.6835	2000	3.9886	0.9985	0.9931
5.9809	2.1044	2500	2.6265	0.9988	0.6725
1.9546	2.5253	3000	0.6612	0.4666	0.1509
1.5891	2.9461	3500	0.5417	0.3780	0.1210
1.4711	3.3670	4000	0.4786	0.3412	0.1086
1.4025	3.7879	4500	0.4494	0.3345	0.1054
1.4410	4.2088	5000	0.4487	0.3194	0.1019
1.2083	4.6296	5500	0.4384	0.3156	0.1002
1.1199	5.0505	6000	0.4265	0.3091	0.0986
1.2789	5.4714	6500	0.4436	0.3083	0.0986
1.2491	5.8923	7000	0.4328	0.3088	0.0985
1.2187	6.3131	7500	0.4934	0.3369	0.1062

Framework versions

Transformers 5.0.0
Pytorch 2.10.0+cu128
Datasets 4.0.0
Tokenizers 0.22.2

Citation

bibtex

@article{waxalnet2026,
  title  = {The WAXAL ASR Benchmark: Fine-Tuned Edge Models Across 19 African Languages},
  author = {Olufemi, Victor Tolulope and Babatunde, Oreoluwa and Njema, Ramsey and
             Gbotemi, Bolarinwa and Yen, Wanchi Lucia and Uzodinma, John and
             Ajayi, Sunday and Williams, Oluwademilade and Moshood, Kausar and
             Anyaele, Innocent Elendu and Arefaine, Akebert Tesfahunegn and
             Hunzwi, Candace and Daniel, Wongel Dawit and Namuganga, Emmilly Immaculate and
             Kadima, Cleophas and Bahizire, Athanase Biluge and Ranaivoson, Onitsiky and
             Aaron, Emmanuel and Ladislaus, Nicholaus Dismas and Muhammed, Idris and
             Simenya, Jonathan Enoch and Koome, Martin and Endaylalu, Matewos Tegete and
             Adeyemo, Peter Ifeoluwa and Birindwa, Hondi Prisca and Eze-Mbey, Ukachi Agnes and
             Oduro-Yeboah, Yacoba and Aremu, Toluwani and Adjovi, Pericles and
             Ngueajio, Mikel K and Mitra, Prasenjit},
  year   = {2026},
  note   = {arXiv preprint arXiv:2606.02375}
}

Authors

Victor Tolulope Olufemi · Oreoluwa Babatunde · Ramsey Njema · Bolarinwa Gbotemi · Wanchi Lucia Yen · John Uzodinma · Sunday Ajayi · Oluwademilade Williams · Kausar Moshood · Innocent Elendu Anyaele · Akebert Tesfahunegn Arefaine · Candace Hunzwi · Wongel Dawit Daniel · Emmilly Immaculate Namuganga · Cleophas Kadima · Athanase Biluge Bahizire · Onitsiky Ranaivoson · Emmanuel Aaron · Nicholaus Dismas Ladislaus · Idris Muhammed · Jonathan Enoch Simenya · Martin Koome · Matewos Tegete Endaylalu · Peter Ifeoluwa Adeyemo · Hondi Prisca Birindwa · Ukachi Agnes Eze-Mbey · Yacoba Oduro-Yeboah · Toluwani Aremu · Pericles Adjovi · Mikel K Ngueajio · Prasenjit Mitra

Acknowledgements

We thank the following contributors for their language expertise and native-speaker evaluation support: Ajara Oyinloye, Abubakari Sadic Mohammed, Hafiz Adjei, Aliga Norah Lele, Marie-Louise B. Ndamuso, and Odong Diana.

This work was supported by Lynguallabs (compute, researchers & storage), Open Token (compute resources), and CMU Africa (researchers & native speakers).