Datasets:
Limitations
Read this before quoting anything here.
1. One domain, and it is the easy one
Every number is measured on people reading prompts aloud into a phone. One speaker, close microphone, scripted grammatical sentence, no overlap, no music.
That is not how anyone talks.
We measured the gap on spontaneous podcast audio using teacher agreement as a proxy: the same models disagree with each other roughly three times more on spontaneous speech than on read prompts (mean pairwise distance 0.32 against 0.10). We could not put an absolute error rate on it because we had no ground truth for spontaneous speech.
Do not quote these figures as general Amharic ASR performance. They describe one domain, and it is the domain most favourable to every model in the table.
2. Greedy decoding only
No beam search, no language model in the base numbers, no hyperparameter search per model. Every figure is a lower bound on what the model can do.
This limitation is not evenly distributed. Sequence-to-sequence models are usually deployed with beam search and gain more from it than CTC models typically do. The within-family comparisons are sound; the cross-family comparison should be treated as provisional.
3. Normalisation is a choice, and it moves the numbers
Folding Ge'ez homophone families is defensible, since those characters are not distinguished in modern pronunciation, but it is a choice, and a stricter scorer would produce higher error rates for everyone. The script is published so the choice is inspectable rather than implicit. Numbers from a different normaliser are not comparable to these.
4. Small by benchmark standards
1,548 clips and 4.72 hours. Enough to separate models that differ by more than about half a point of CER, as the confidence intervals show, and not enough to rank models that are genuinely close. Where the paired test says a difference is not distinguishable, treat it as a tie however different the point estimates look.
5. Single run, no seed variance
Each model was evaluated once. Greedy decoding is deterministic, so there is no sampling variance to average over, but nor is there any check against a transient environment problem. The full inputs and outputs are published so that an independent rerun is cheap.
6. What is not measured at all
- Real-time factor, latency, memory. A model that is slightly better and ten times slower is not better for most deployments.
- Robustness to noise, telephone bandwidth, or code-switching into English, which is extremely common in Ethiopian speech.
- Speaker demographics beyond what the corpus records. If a model is worse for women, or for particular regional accents, nothing here would reveal it.
That last one is a real gap and we intend to close it, since the corpus carries gender, age band and region for every clip.
7. The language model damages code-switched speech
Found by talking to the demo rather than by any measurement here.
Recording a sentence that mixed English into Amharic, the way people actually speak in Addis, the speech model heard "my name is" and transliterated it phonetically into Ge'ez, which was correct. The language model then changed it into something else.
It was confidently correcting a word that was already right. This is not a bug, it is the language model being what it is: trained on 12.8 million lines of Amharic text, in which transliterated English never appears. Seeing one, it concludes the speech model must have erred and pulls toward something more Amharic-shaped.
The more English a clip contains, the worse that trade becomes. Since English code-switching is extremely common in Ethiopian speech, this is a real limitation and not an edge case.
Nothing in this benchmark could have caught it. The test sentences come from news text and contain almost no English. It took one person speaking naturally for thirty seconds.
The fix is to train the language model on Amharic as it is actually spoken, English included, which requires transcripts of spontaneous speech. That is the same missing ingredient as limitation 1.