snapwre/amharic-asr-benchmark

Amharic ASR Benchmark

An evaluation of open speech recognition models for Amharic, on a test set with certain labels and honest statistics.

16 models. 1,548 clips. 4.72 hours. Every hypothesis published.

Published by Dataset.ET.

Read this table first

Round 1 of this benchmark rested on a single clean claim: every model predated our dataset, so none could have trained on it. That claim no longer holds. Models trained on snapwre/amharic-speech now exist, and others are trained on data their authors do not disclose.

Mixing those into one ranking would make the leaderboard meaningless. So results are split by what the model was allowed to see, and the three tables are not comparable to each other.

1. Clean

Published before this test set existed, or verifiably not trained on it. These are the numbers to quote when you want a fair comparison.

model	CER	95% CI	WER	speed
boazsew/Ethio-ASR-w2v-bert-2.0-uf	0.0924	[0.0872, 0.0975]	0.2873	154x	CTC, w2v-BERT 2.0, WAXAL only
badrex/Ethio-ASR-amharic	0.0946	[0.0896, 0.0999]	0.2845	141x	CTC, monolingual Amharic, 606M
badrex/Ethio-ASR-multilingual-600M	0.0991	[0.0939, 0.1042]	0.2984	133x	CTC, w2v-BERT 2.0, five languages, 606M
b1n1yam/shook-medium-amharic-2k	0.1147	[0.1087, 0.1202]	0.2943	24x	Seq2seq, Whisper medium, 764M
aadel4/omniASR-CTC-1B-v2	0.1231	[0.1181, 0.1280]	0.4627	384x	CTC, Meta Omnilingual ASR, 1,600 languages
badrex/Ethio-ASR-multilingual-1B	0.1308	[0.1249, 0.1366]	0.3894	268x	CTC, MMS-based, 963M
badrex/Ethio-ASR-multilingual-94M	0.1538	[0.1477, 0.1596]	0.4622	667x	CTC, smallest of the family
badrex/Ethio-ASR-multilingual-300M	0.1590	[0.1529, 0.1646]	0.4702	531x	CTC
boazsew/Ethio-ASR-afrihubert-uf	0.1653	[0.1588, 0.1714]	0.4916	664x	CTC, AfriHuBERT encoder
facebook/seamless-m4t-v2-large	0.2306	[0.2202, 0.2421]	1.0334	55x	Seq2seq, lists amh as supported
agkphysics/wav2vec2-large-xlsr-53-amharic	0.2500	[0.2446, 0.2554]	1.1560	535x	CTC, XLSR-53
b1n1yam/shook-tiny-amharic-stage2-polish	0.2735	[0.2648, 0.2820]	0.6200	70x	Seq2seq, 37.8M
openai/whisper-large-v3	1.5105	[1.4879, 1.5342]	1.3837	14x	Control, no Amharic training

The top two are a statistical tie. Paired bootstrap gives p = 0.061 with a confidence interval that crosses zero. Ethio-ASR-amharic is fractionally better on WER, Ethio-ASR-w2v-bert-2.0-uf fractionally better on CER. Pick either.

2. Trained on this dataset

Uses snapwre/amharic-speech in training. Legitimate, disclosed, and not comparable to the table above, because the test set is drawn from the same corpus as its training data.

model	CER	95% CI	WER	speed	trained on
boazsew/Ethio-ASR-w2v-bert-2.0-combined-am-uf	0.0576	[0.0534, 0.0619]	0.1852	154x	WAXAL-am 189.7h + snapwre 18.3h

3. Undisclosed training data

Published after this test set, with training data the model card does not enumerate. We cannot verify contamination either way.

model	CER	95% CI	WER	speed
b1n1yam/qwen3-asr-0.6b-amharic-mixed	0.1837	[0.1759, 0.1910]	0.4666	12.7x	Qwen3-ASR-0.6B fine-tune
b1n1yam/qwen3-asr-0.6b-amharic-gold-silver	0.1857	[0.1775, 0.1942]	0.4765	12.8x	Sibling, gold plus silver labels

These two are a statistical tie, p = 0.203. Whatever "gold plus silver" changed, it did not change accuracy here.

The finding

We tested a new architecture family, two large multilingual systems from Meta, the most-downloaded Amharic ASR model on the Hub, and three new w2v-BERT variants. Not one of them beat what was already the best clean model.

The only result that moved was a model trained on this dataset, and it moved a long way: 0.1852 WER against 0.2845, a 35% relative reduction.

Architecture is not the bottleneck for Amharic ASR right now. Data is.

An independent study, reproduced

boazsew/Ethio-ASR-w2v-bert-2.0-combined-am-uf comes from a University of Florida technical report by Boaz Tulu, Does Mixing snapwre with WAXAL Improve Amharic ASR? A matched-step, cross-domain study with w2v-BERT-2.0. Two models, same architecture, same recipe, same step budget, differing only in training data.

That paper reports a 10.25 point WER improvement on this test set from adding our 18.3 hours. Measured here independently, with our own harness and normaliser: 28.73 to 18.52, a 10.21 point gap.

Four hundredths of a point apart. We did not coordinate on this and had no involvement in that study.

The paper also reports the result that matters more: adding this data improved performance on WAXAL's own test set too, 25.28 to 24.71. That is not domain adaptation, that is the data carrying information the larger corpus did not have.

The most-downloaded Amharic ASR model does not work

agkphysics/wav2vec2-large-xlsr-53-amharic has 148,577 downloads, more than any other Amharic-specific ASR model on the Hub. It scores 1.1560 WER.

Above 1.0 means it produces more errors than there are words in the reference. If you picked a model by sorting on downloads, this is what you got.

Meta's multilingual models split

omniASR-CTC-1B-v2, the Omnilingual ASR release covering 1,600 languages, lands at 0.1231 CER. Respectable for a model that treats Amharic as one language among sixteen hundred, and better than three purpose-built Amharic models here.

seamless-m4t-v2-large lists amh as a supported language and scores 1.0334 WER, joining whisper-large-v3 and agkphysics above the 1.0 line. Listing a language and recognising it are different claims.

Are the differences real?

Every model is scored on the same 1,548 clips, so comparing marginal confidence intervals is the wrong test. They can overlap while the difference is real, and they can separate while it is not. A paired bootstrap over 2,000 resamples of clips, on adjacent pairs:

comparison	mean CER difference	95% CI	distinguishable
combined-am-uf vs w2v-bert-2.0-uf	-0.0348	[-0.0376, -0.0322]	yes
w2v-bert-2.0-uf vs Ethio-ASR-amharic	-0.0022	[-0.0048, +0.0007]	no, p = 0.061
Ethio-ASR-amharic vs multilingual-600M	-0.0045	[-0.0067, -0.0022]	yes
multilingual-600M vs shook-medium	-0.0155	[-0.0194, -0.0120]	yes
shook-medium vs omniASR-CTC-1B-v2	-0.0085	[-0.0123, -0.0044]	yes
omniASR-CTC-1B-v2 vs multilingual-1B	-0.0077	[-0.0111, -0.0040]	yes
multilingual-1B vs multilingual-94M	-0.0230	[-0.0263, -0.0200]	yes
multilingual-94M vs multilingual-300M	-0.0052	[-0.0080, -0.0023]	yes
multilingual-300M vs afrihubert-uf	-0.0062	[-0.0098, -0.0026]	yes
afrihubert-uf vs qwen3-asr-mixed	-0.0184	[-0.0239, -0.0134]	yes
qwen3-asr-mixed vs qwen3-asr-gold-silver	-0.0022	[-0.0077, +0.0027]	no, p = 0.203
qwen3-asr-gold-silver vs seamless-m4t-v2	-0.0448	[-0.0562, -0.0343]	yes
seamless-m4t-v2 vs xlsr-53-amharic	-0.0195	[-0.0279, -0.0097]	yes
xlsr-53-amharic vs shook-tiny	-0.0235	[-0.0299, -0.0174]	yes
shook-tiny vs whisper-large-v3	-1.2373	[-1.2627, -1.2117]	yes

Thirteen of fifteen adjacent pairs are separable. The two that are not are reported as ties above rather than ranked.

The harness reproduced itself exactly

Three round 1 models were re-run in round 2 for exactly this reason. Not approximately, identically:

model	round 1	round 2
badrex/Ethio-ASR-amharic	0.0946 / 0.2845	0.0946 / 0.2845
badrex/Ethio-ASR-multilingual-600M	0.0991 / 0.2984	0.0991 / 0.2984
b1n1yam/shook-medium-amharic-2k	0.1147 / 0.2943	0.1147 / 0.2943

Round 1 published only summary numbers, not per-clip hypotheses, which is why those three had to be re-run before rounds could be compared at all. Round 2 publishes every hypothesis. Round 3 will not have this problem.

One model failed

mintesnotfikir/whisper-medium-amharic failed to load its processor (TypeError: expected str, bytes or os.PathLike object, not NoneType) and was skipped by the smoke test after one second. Not scored, not ranked, not hidden.

Whisper large v3 is a control, not a contender

Its CER of 1.5105 is above 1.0, meaning it emits more wrong characters than the reference contains. It is not transcribing Amharic badly, it is hallucinating. This matches the Ethio-ASR paper's report of over 100% WER on Ethiopian languages, and it is included so the scale of the other numbers is legible.

On word error rate

The best clean model sits at 28.45% WER. Character rate is the headline for an agglutinative language, but if you are comparing against word-rate figures published elsewhere, that is the number. No clean open model we tested is close to the 15% WER that would make Amharic recognition feel solved.

Does an Amharic language model help?

A CTC model decides each audio frame independently and has no idea what an Amharic word is, so when the audio is ambiguous it emits letters that spell nothing. An n-gram language model rescores candidates during beam search and pulls the output toward real words.

We trained a 5-gram on 12,868,562 lines of open Amharic text, with every test and validation sentence removed. Alpha and beta were tuned on the validation split and applied unchanged to test.

model	WER greedy	WER + LM	change	CER greedy	CER + LM	change
Ethio-ASR-amharic	0.2845	0.2483	-12.7%	0.0946	0.0965	+2.0%
Ethio-ASR-multilingual-600M	0.2984	0.2566	-14.0%	0.0991	0.1034	+4.3%
Ethio-ASR-multilingual-1B	0.3894	0.3154	-19.0%	0.1308	0.1370	+4.8%

Word error rate improves 13 to 19% relative. Character error rate gets slightly worse. That is not a contradiction, it is the language model working as designed. It pulls output toward real words, so whole words match far more often. When it settles on the wrong real word it changes several characters at once, so character error ticks up slightly. For anything a person reads, the word rate is what matters.

The best result on this benchmark is now 24.83% WER, from Ethio-ASR-amharic with the language model, against 28.45% without it.

Two caveats. The best alpha and beta both landed on the edge of the search grid, so a wider search would probably find a better setting. And the trained model is 17 GB unpruned, which is impractical to distribute; a pruned version is the obvious next step.

Decontamination

Removing test sentences from 12.8 million lines of text is where a language model quietly cheats. Exact whole-line matching found zero contaminating lines. Substring matching over all 2,990 held-out sentences found four, buried inside longer paragraphs. Small, but they were there, and nothing would have reported an error if we had missed them.

Pruned, so you can actually use it

The unpruned 5-gram is 17 GB. Nobody downloads that, and it will not fit in a demo. Pruning drops the n-grams seen once or twice at higher orders, which are mostly noise, and quantisation stores probabilities in 8 bits instead of 32.

Decoded with badrex/Ethio-ASR-amharic on the full test set. The greedy baseline for comparison is CER 0.0946, WER 0.2845.

file	size	CER	WER	WER change
am-5gram-light.bin	1839 MB	0.0968	0.2486	-12.6%
am-5gram-medium.bin	988 MB	0.0969	0.2494	-12.3%
am-5gram-aggressive.bin	538 MB	0.0973	0.2506	-11.9%

The 538 MB model does what the 17 GB model does. Thirty-two times smaller for a difference of 0.002 WER, which is noise at this sample size. Use kenlm/am-5gram-aggressive.bin unless you have a reason not to.

from huggingface_hub import hf_hub_download
from pyctcdecode import build_ctcdecoder

lm = hf_hub_download("snapwre/amharic-asr-benchmark",
                     "kenlm/am-5gram-aggressive.bin", repo_type="dataset")
decoder = build_ctcdecoder(labels, lm, alpha=0.3, beta=1.5)

Or try it without installing anything: huggingface.co/spaces/Chapimenge/amharic-asr-demo

Does beam search change the ranking?

Greedy decoding is not neutral between model families, so this checks whether it disadvantaged the sequence-to-sequence models.

model	CER at beam width 5
Ethio-ASR-amharic	0.0958
Ethio-ASR-multilingual-600M	0.0996
Ethio-ASR-multilingual-1B	0.1306
shook-medium-amharic-2k	0.1127
shook-tiny-amharic-stage2-polish	0.2453

It does not change the ranking. Against the greedy figures the CTC models move by at most 0.0012 CER, and two of the three get very slightly worse, which is expected: with no language model to score candidates a beam has nothing to prefer. Whisper gains a little and pays heavily in speed, 714 seconds to 4,151 for the medium model.

So greedy decoding was not quietly disadvantaging either family, and the ranking above stands as measured.

Speed

Measured on the same run, one A100, batch size 16. This is throughput for bulk transcription, not single-clip latency.

model	x realtime	CER
Ethio-ASR-multilingual-1B	268x	0.1308
Ethio-ASR-amharic	141x	0.0946
Ethio-ASR-multilingual-600M	133x	0.0991
shook-tiny-amharic-stage2-polish	70x	0.2735
shook-medium-amharic-2k	24x	0.1147
whisper-large-v3	14x

Two things fall out of this that accuracy alone hides.

The most accurate model is also six times faster than the next most accurate. There is no accuracy-for-speed trade to make at the top of this table.

A 37.8M parameter model is slower than a 606M one. Architecture dominates size: a CTC model does a single forward pass over the audio, while a sequence-to-sequence model decodes one token at a time and pays for every character it emits. If throughput matters, that difference is larger than anything parameter count will tell you.

What makes this worth trusting

The labels are certain. A contributor was shown a sentence and read it aloud, so the reference existed before the audio. Most speech benchmarks score against transcripts someone typed while listening, which carries an unmeasured error rate of its own.

Contamination is stated, not assumed. Round 1 could claim every model predated the test set. Round 2 cannot, so each model is placed in one of three categories by what its authors disclose:

category	basis
clean	published before 2026-08-25, or training data enumerated and ours absent
trained on this dataset	the model card names snapwre/amharic-speech
undisclosed	published after 2026-08-25, training data not enumerated

The third category is not an accusation. It is the honest label for a model whose card says only that it used "a privately maintained corpus". We cannot check, so we do not rank it against models we can.

Speaker-disjoint splits. No voice appears in more than one split.

The outputs are published. Every model's raw transcript for all 1,548 clips is in the repo. Do not trust these numbers. Recompute them.

What this does not tell you

This is one domain: people reading prompts into a phone. On spontaneous speech the same models are far worse. We measured roughly three times the disagreement between them on podcast audio. These are not general Amharic ASR figures. See docs/LIMITATIONS.md.

The test set is public, which is why the category split above had to exist at all. Every model published from here on may have seen it. An uncontaminated successor with withheld references is the next piece of work, and it is needed before this benchmark ages out.

Decoding is greedy, so every number is a lower bound, and that penalty is not evenly distributed between CTC and sequence-to-sequence models.

What went wrong

docs/FAILURES.md is the honest engineering log: failures including three that produced results which looked correct, a health check that reported a working system as broken, and a shutdown timer we triggered by trying to disable it.

Round 2 added two more. A dependency added for one model upgraded numpy to 2.x, which broke every system package compiled against 1.x and killed the box on preflight ten minutes in. And a launcher edit searched for a heredoc terminator from the start of the file, matched the opening delimiter instead of the closing one, produced an empty slice, and a replace("") inserted new text between every character of a 22 KB file.

It is the most useful document here. Benchmarks are usually published as though they fell out of the sky.

Contents

README.md                          this
docs/METHOD.md                     test set, scoring, statistics
docs/FAILURES.md                   what went wrong and what it cost
docs/LIMITATIONS.md                read before quoting
results/phase1_greedy.json         round 1 numbers
results/bootstrap.json             intervals and paired tests
hypotheses/                        round 1 raw output, per clip
runs/<run-id>/results.json         round 2 numbers
runs/<run-id>/hyps/                round 2 raw output, per clip
runs/<run-id>/references.json      the references, so anything can be rescored
kenlm/                             pruned Amharic 5-gram language models
scripts/                           the evaluation code

Reproducing

python scripts/bench_asr.py    --push-to <your-repo>   # round 1
python scripts/bench_round2.py --push-to <your-repo>   # round 2
python scripts/bootstrap_ci.py build/am-v0.2.0 hypotheses/

The test set is snapwre/amharic-speech.

Citing

@misc{datasetet2026amharicasr,
  title  = {Amharic ASR Benchmark: an evaluation of open speech models},
  author = {Dataset.ET},
  year   = {2026},
  url    = {https://huggingface.co/datasets/snapwre/amharic-asr-benchmark}
}

If you are citing the finding that adding this dataset improves Amharic ASR, cite the study that established it, not us:

@techreport{tulu2026combined,
  author      = {Boaz Tulu},
  title       = {Does Mixing snapwre with WAXAL Improve Amharic ASR?
                 A matched-step, cross-domain study with w2v-BERT-2.0},
  institution = {University of Florida},
  year        = {2026}
}

Contact

Corrections and additional models are welcome. Open a discussion. If you built one of these models and think a decoding choice disadvantaged it, say so and we will rerun it.

If your model is in the undisclosed category and you would like it moved, tell us what it trained on and we will move it.

Chapi, Dataset.ET