cs.CLMar 24, 2026
SaveEthio-ASR: Joint Multilingual Speech Recognition and Language Identification for Ethiopian Languages
Organizations: Saarland University, Germany · University College London, UK · Ethiopian AI Institute, Ethiopia · HiLCoE, Ethiopia ·
Abstract
We present Ethio-ASR, a suite of multilingual CTC-based automatic speech recognition (ASR) models jointly trained on five Ethiopian languages: Amharic, Tigrinya, Oromo, Sidaama, and Wolaytta. These languages belong to the Semitic, Cushitic, and Omotic branches of the Afroasiatic family, and remain severely underrepresented in speech technology despite being spoken by the vast majority of Ethiopia's population. We train our models on the recently released WAXAL corpus using several pre-trained speech encoders and evaluate against strong multilingual baselines, including OmniASR. Our best model achieves an average WER of 30.48% on the WAXAL test set, outperforming the best OmniASR model with substantially fewer parameters. We further provide a comprehensive analysis of gender bias, the contribution of vowel length and consonant gemination to ASR errors, and the training dynamics of multilingual CTC models. Our models and codebase are publicly available to the research community.
Figures & tables
\fontspec_if_language:nTF ENG \addfontfeature Language=EnglishFigure 1: Average WER (%) versus model size, where lower WER indicates better performance. Our Ethio-ASR models consistently achieve lower WER across all sizes, outperforming CTC and LLM-based OmniASR baselines.
| Language | ISO | Text sample | Script |
|---|---|---|---|
| English | \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishENG | All human beings are born free and equal in dignity and rights. | Latin |
| Amharic | \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishAMH | \ethiopicfont የሰው ልጆች ሁሉ ነጻ፥ በክብርም በመብትም እኩል ሆነው ተወልደዋል። | Ethiopic |
| Oromo | \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishORM | Namooti hundinuu birmaduu ta’anii mirgaa fi ulfinaanis wal-qixxee ta’anii dhalatan. | Latin |
| Tigrinya | \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishTIR | \ethiopicfont ብመንፅር ክብርን መሰልን ኩሎም ሰባት እንትውለዱ ነፃን ማዕሪን እዮም። | Ethiopic |
| Sidaama | \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishSID | Manchi beetti kalaqamunni wolaphinoho. Ayirrinyunninna qoossotennino taaloho. | Latin |
| Wolaitta | \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishWAL | Ubba asaykka la’an daanawu yelettiis, qassikka bonchchuwaaninne maatan lagge. | Latin |
\fontspec_if_language:nTF ENG \addfontfeature Language=EnglishTable 1: UDHR Article 1 across five Ethiopian languages in this paper with English translation as reference.
| \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishAMH | \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishORM | \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishTIR | \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishSID | \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishWAL | ||
| Train | ||||||
| M | 105.88 | 85.14 | 83.20 | 64.70 | 76.56 | |
| F | 83.83 | 104.75 | 98.67 | 127.53 | 120.75 | |
| All | 189.71 | 189.89 | 181.87 | 192.23 | 197.32 | |
| Validation | ||||||
| M | 6.56 | 5.04 | 6.70 | 4.75 | 0.02 | |
\fontspec_if_language:nTF ENG \addfontfeature Language=EnglishTable 2: Audio duration (hours) by language, split, and gender (M: male voices, F: female voices).
\fontspec_if_language:nTF ENG \addfontfeature Language=EnglishFigure 2: Top: Vocabulary growth curves for Ethiopian languages (WAXAL) and English/French (Multilingual LibriSpeech) across corpus sizes up to 1.4M tokens. Bottom: Type-token ratio (TTR) at 800k tokens.
| Model | # params | Decoder | Amharic | Tigrinya | Oromo | Wolaytta | Sidaama | Avg. |
| OpenAI ASR | ||||||||
| whisper-small | 300M | LM | 157.85 | 174.31 | 152.99 | 184.67 | 182.32 | 170.43 |
| whisper-medium | 786M | LM | 195.05 | 198.06 | 192.64 | 247.30 | 232.91 | 213.19 |
| whisper-large-v3 | 1.6B | LM | 153.03 | 166.20 | 128.65 | 151.04 | 144.08 | 148.60 |
| Meta ASR models | ||||||||
| seamless-m4t-v2-large | 2B | LM | 103.75 | 100.00 | 100.00 | 100.00 | 100.00 | 100.75 |
\fontspec_if_language:nTF ENG \addfontfeature Language=EnglishTable 3: WER (%, lower is better) on the test split of the WAXAL dataset. Bold indicates the best result per language among multilingual Ethio-ASR models, while underlined indicates best performance among all models.
| Model | \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishAMH | \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishORM |
|---|---|---|
| MMS | ||
| mms-1b-all | 29.78 | 64.71 |
| OmniASR | ||
| omniASR-ctc-300m-v2 | 34.41 | 77.53 |
| omniASR-ctc-1b-v2 | 48.23 | 68.38 |
| omniASR-ctc-3b-v2 | 20.59 | 62.76 |
\fontspec_if_language:nTF ENG \addfontfeature Language=EnglishTable 4: WER (%, lower is better) on FLEURS. Note that FLEURS data was included in the pre-training or fine-tuning data of all baseline models but was not seen by our models during training, making this a zero-shot out-of-domain evaluation for Ethio-ASR.
| Model | # params | Acc. |
|---|---|---|
| Ethio-ASR ( \fontspec_if_language:nTF ENG \addfontfeature Language=Englishafrihubert ) | 94M | 99.92 |
| Ethio-ASR ( \fontspec_if_language:nTF ENG \addfontfeature Language=Englishmms-300m ) | 300M | 99.92 |
| Ethio-ASR ( \fontspec_if_language:nTF ENG \addfontfeature Language=Englishmms-1b ) | 1B | 99.91 |
| Ethio-ASR ( \fontspec_if_language:nTF ENG \addfontfeature Language=Englishw2v-bert-2.0 ) | 600M | 99.83 |
\fontspec_if_language:nTF ENG \addfontfeature Language=EnglishTable 5: Language identification (LID) in accuracy (%).
| Model | +LID | –LID |
|---|---|---|
| Ethio-ASR ( \fontspec_if_language:nTF ENG \addfontfeature Language=Englishafrihubert ) | 37.93 1.32 | 38.04 1.23 |
| Ethio-ASR ( \fontspec_if_language:nTF ENG \addfontfeature Language=Englishmms-300m ) | 36.77 1.21 | 36.15 1.20 |
| Ethio-ASR ( \fontspec_if_language:nTF ENG \addfontfeature Language=Englishmms-1b ) | 33.94 1.24 | 33.86 1.31 |
| Ethio-ASR ( \fontspec_if_language:nTF ENG \addfontfeature Language=Englishw2v-bert-2.0 ) | 33.02 1.28 | 33.28 1.28 |
\fontspec_if_language:nTF ENG \addfontfeature Language=EnglishTable 6: WER (%) under +LID and –LID conditions on the WAXAL test set. Values shown as mean half-width of the 95% bootstrap CI. No difference is statistically significant ( ).
| \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishAMH | \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishORM | \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishSID | \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishTIR | |
|---|---|---|---|---|
| Ethio-ASR ( \fontspec_if_language:nTF ENG \addfontfeature Language=Englishafrihubert ) | ||||
| Male | 32.08 | 28.34 | 29.96 | 56.78 |
| Female | 30.02 | 25.11 | 33.57 | 38.24 |
| +2.06 | +3.23 | 3.61 | +18.54 | |
| Ethio-ASR ( \fontspec_if_language:nTF ENG \addfontfeature Language=Englishmms-300m ) | ||||
| Male | 32.81 | 29.17 | 31.63 | 58.14 |
\fontspec_if_language:nTF ENG \addfontfeature Language=EnglishTable 7: WER (%) by gender across languages (excluding Wolaytta). = Male Female; positive values indicate higher error for male speakers.
\fontspec_if_language:nTF ENG \addfontfeature Language=EnglishFigure 3: Training dynamics of Ethio-ASR with \fontspec_if_language:nTF ENG \addfontfeature Language=Englishw2v-BERT-2.0 and \fontspec_if_language:nTF ENG \addfontfeature Language=EnglishMMS-1B as pre-trained models. This analysis reveals that both models learn to transcribe Latin-based scripts before Ge’ez script.
\fontspec_if_language:nTF ENG \addfontfeature Language=EnglishFigure 4: Effect of vowel length and geminate normalization on WER across Oromo, Sidaama, and Wolaytta. Results are shown for four conditions: no normalization, vowel length normalization, geminate normalization, and both normalized.
More in this topic
Speech Processing
Speech processing develops computational methods for recognizing, synthesizing, analyzing, and transforming spoken audio.
See all 400 papers