Automatic speech recognition (ASR) remains challenging for low-resource languages, where the scarcity of transcribed speech limits the development of accurate systems. This challenge is particularly acute for African languages such as Ewe, which has millions of speakers but remains underrepresented in publicly available ASR resources and benchmarks. In this work, we consolidate, normalize, and partition existing Ewe speech resources into a standardized 174-hour Ewe ASR benchmark, the largest to date to our knowledge. We fine-tune and evaluate pretrained speech models, including Whisper and AfriHuBERT, to establish strong baselines on our benchmark. Whisper-medium achieves a word error rate (WER) of 37.98%, while AfriHuBERT achieves 39.70% on the test set. We further provide a qualitative analysis of common recognition errors. We release the benchmark under the name Nufo, derived from the Ewe word for "talk", accompanied by models and evaluation resources, to facilitate future research on Ewe and other underrepresented African languages.
Figures & tables
Source
Train
Valid.
Test
WAXAL
16,078
1,614
1,443
WAXAL + BibleTTS
36,011
1,614
1,443
WAXAL + BibleTTS + MINA
52,417
1,614
1,443
Table 1: Training, validation, and test set sizes, measured in number of utterances. Validation and test sets are fixed across all configurations to compare performance with the addition of more training data. Data comes from WAXAL [ 3 ] , BibleTTS [ 7 ] , and Common Voice - Mina (MINA) [ 2 ] .
Whisper
AfriHuBERT
medium
base
Parameters
769 M
74 M
97.4 M ⋆
Effective epochs
6.41
10.07
27
Batch / accum.
4 / 4
4 / 4
2 / –
Lang. condition
yo
yo
–
Objective
Seq2Seq
Seq2Seq
CTC
Table 2: Models and fine-tuning settings.
Data
Model
WER
CER
full
np
np_nd
full
np_nd
W
Whisper-base
46.84
42.10
39.36
16.23
14.18
Whisper-med
39.69
35.49
32.65
13.03
11.20
W+B
Whisper-base
44.92
40.48
37.73
15.29
13.22
Whisper-med
38.71
34.32
31.35
12.48
10.50
W+B+M
Whisper-base
44.72
40.28
37.57
15.07
13.08
Table 3: ASR results on the fixed 1,443-utterance test set (%). W = WAXAL, B = BibleTTS, M = Common Voice - Mina. Columns np , np_nd , represent performance when punctuation is removed and when punctuation and diacritics are removed, respectively.
Reference
Hypothesis
Error type / analysis
Whisper-medium
…utsu deka le mega foni e asi …
…utsu deka lé megafoni e asi …
Word boundary: mega foni → megafoni .
ame geewo tsi atsitre e goto.
ame geewo tsatsitre e goto.
Word boundary: tsi atsitre → tsatsitre .
…i kuku fitaa eye …
…kuku fita eye …
Ortho: fitaa → fita ; epenthetic vowel deletion.
…a Gego violet kple a Gego i …
…a Gego baulet kple a Gego i …
Lexical substitution: violet → baulet .
AfriHuBERT-CTC
Table 4: Representative errors on the Ewe test set. Ortho denotes orthography errors.