WAXAL: A Large-Scale Multilingual African Language Speech Corpus
Authors: Abdoulaye Diack, Perry Nelson, MohamedElfatih MohamedKhair, Emmanuel Asiedu Brempong, Subhashini Venugopalan, Tavonga Siyavora, Vusumuzi Dube, Kwaku Agbesi, Angela Nakalembe, Jason Hickey, Uche Okonkwo, Abhishek Bapna, Isaac Wiafe, Raynard Dodzi Helegah, Elikem Doe Atsakpo, Charles Nutrokpor, Fiifi Baffoe Payin Winful, Kafui Kwashie Solaga, Jamal-Deen Abdulai, Akon Obu Ekpezu, Audace Niyonkuru, Samuel Rutunda, Boris Ishimwe, Michael Melese, Engineer Bainomugisha, Joyce Nakatumba-Nabende, Andrew Katumba, Claire Babirye, Jonathan Mukiibi, Vincent Kimani, Samuel Kibacia, James Maina, Fridah Emmah, Ahmed Ibrahim Shekarau, Ibrahim Shehu Adamu, Yusuf Abdullahi, Howard Lakougna, Kode Niane, Bob MacDonald, Pooja Rao, Hadar Shemtov, Prince Mireku, Aisha Walcott-Bryant, Moustapha Cisse, Avinatan Hassidim, Jeff Dean, Yossi Matias
Recent advances in speech technology predominantly favor high-resource languages, leaving speakers of Sub-Saharan African languages underserved. To bridge this divide, we introduce WAXAL, an open, large-scale speech dataset spanning 27 African languages representing over 100 million speakers. WAXAL comprises over 14,100 hours of speech collected in partnership with African academic and community organizations to build local research capacity: an Automatic Speech Recognition (ASR) corpus with 2,340 hours of transcribed, image-prompted natural speech across 19 languages (plus ~11,760 untranscribed hours) and a studio-quality Text-to-Speech (TTS) corpus of ~247 hours across 15 language configurations. We empirically benchmark three architecturally distinct ASR models -- Gemma 3n 2B, Whisper-Large-v3, and MMS-1B-all -- across the 19 ASR languages on strictly speaker-disjoint splits. Fine-tuning on WAXAL yields substantial gains, reducing macro-average WER by up to 58% for Whisper-Large-v3, 56% for Gemma 3n 2B, and 17% for MMS-1B-all (and Character Error Rate by up to 70%). WAXAL-fine-tuned models generalize zero-shot to three out-of-distribution benchmarks: FLEURS, Mozilla Common Voice, and Sunbird SALT. Finally, an LLM-as-judge semantic evaluation confirms that fine-tuning preserves speaker intent beyond surface-level transcription metrics. WAXAL is released at https://huggingface.co/datasets/google/WaxalNLP.
Figures & tables
Figure 1: Examples of some of the images used as prompts to elicit natural speech for the ASR dataset.
Figure 2: WAXAL-ASR: Distribution of the transcribed and total recordings per language.
Dataset
# Langs
# African
ASR Hrs
Avg Hrs/Lang
Elicitation
Domain
FLEURS [ 5 ]
102
∼ 20
∼ 1,200
∼ 12
Scripted reading
Wikipedia/News
Common Voice [ 3 ]
290
80+
19,000+
Skewed
Crowdsourced reading
Open/Varied
MMS-Lab [ 15 ]
1,107
300+
500,000+
Skewed
Scripted reading
Religious text
Afri-voice [ 6 ]
6
6
600
100
Image-prompted & Scripted
Open-domain
Afrivox [ 11 ]
18
16
∼ 64
0.01–6
—
Medical
WAXAL (ours)
27
27
∼ 2,340
∼ 30–220
Image-Prompted, Spontaneous
Open/Varied
Table 1: WAXAL has spontaneous elicitation and uniform scale vs. related African speech datasets.
Figure 3: Demographics of ASR participants and Per-Language Distribution of TTS data.
Dataset
# Languages
# Speakers
Total Hours
# Utterances
WAXAL-TTS
15
103
∼ 247
25,536
WAXAL-ASR (Transcribed)
19
5,021
∼ 2,340
438,229
WAXAL-ASR (All)
19
5,021+
14,140+
∼ 1,956,877
Table 2: Statistics of the WAXAL TTS and ASR datasets.
Figure 4: Zero-shot vs. fine-tuned Word Error Rate (WER) and Character Error Rate (CER) per language on the speaker-disjoint WAXAL test sets ( v2 ) across three models. Teal lines ( ∙ ) show Zero-Shot error rates; coral lines ( ) show Fine-Tuned error rates ( Mean±95% CI over N≥3 runs per language). Fine-tuning consistently and significantly reduces both WER and CER across models and languages. Full numerical tables with 95% CIs are provided in Table 4 (WER) and Table 5 (CER) in Appendix A .
Figure 5: Out-of-Distribution (OOD) Generalization across Three External Benchmarks: FLEURS v2 (6 languages), Mozilla Common Voice (4 languages), and Sunbird SALT (3 East African languages). Models are fine-tuned solely on speaker-disjoint WAXAL ( v2 ) splits without exposure to target benchmark data. Note that FLEURS is in-domain for zero-shot MMS-1B-all , whereas on truly external benchmarks ( Common Voice ), WAXAL fine-tuning reduces WER across all three models including MMS-1B-all . Detailed WER and CER are provided in Table 7 in Appendix A .
Figure 6: Gemma 3n 2B (fine-tuned): surface accuracy ( 1−WER ) vs. LLM-as-judge meaning-preservation scores per language. Languages are sorted by Discrete score (high to low). Dashed lines mark the cross-language means: Discrete =0.667 and Continuous =0.744 . Higher is better for all three metrics. Continuous scores consistently exceed Discrete scores ( +0.077 on average), reflecting partial-credit awards that Discrete scoring quantizes away. Languages with lower WER (e.g., Wolaytta, Shona) also rank highest on both semantic rubrics. Detailed Discrete ( ∈{0.0,0.5,1.0} ) and Continuous ( ∈[0.0,1.0] ) scores are shown in Table 10 in Appendix B .
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Language
ISO
# Speakers
Train
Val
Test
Total Utt.
Hours
Acholi ⋆
ach
170
4,108
519
528
5,155
32.3
Oromo †
orm
441
38,185
3,078
3,782
45,045
220.0
Akan †
aka
404
10,107
1,123
1,522
12,752
102.0
Amharic †
amh
432
38,022
2,912
3,420
44,354
218.0
Dagaare ⋆
dga
240
15,071
1,892
1,910
18,873
104.7
Dagbani ⋆
dag
235
14,231
1,750
1,838
17,819
98.5
Appendix
Table 3: Transcribed WAXAL-ASR dataset statistics per language under the strictly speaker-disjoint splits ( v1 for the 6 Voice Donor languages ( † ), v2 for the 13 re-split Makerere and Clear Global languages ( ⋆ ), guaranteeing 0% train–test speaker overlap across all 19 languages).
Gemma-3n-E2B-it
Whisper-large-v3
MMS-1B-all
Language
ISO
Zero-Shot
Fine-Tuned ( ± 95% CI)
Zero-Shot
Fine-Tuned ( ± 95% CI)
Zero-Shot
Fine-Tuned ( ± 95% CI)
Acholi ⋆
ach
0.9398
0.4175 ± 0.0052
1.0552
0.3853 ± 0.0146
0.5059
0.4158 ± 0.0445
Akan †
aka
0.9295
0.3533 ± 0.0385
0.9917
0.3477 ± 0.0254
0.5699
0.3900 ± 0.0074
Amharic †
amh
0.8568
0.4120 ± 0.0903
1.5240
0.9314 ± 0.0365
0.5715
0.4761 ± 0.0388
Dagbani ⋆
dag
0.9784
0.4337 ± 0.0197
1.2057
0.4417 ± 0.0475
—
0.4754 ± 0.0027
Dagaare ⋆
dga
0.9502
0.4456 ± 0.0171
0.9894
0.4066 ± 0.0064
0.5148
0.4133 ± 0.0042
Appendix
Table 4: Word Error Rate (WER, decimal; lower is better) on the speaker-disjoint WAXAL test sets ( v1 for 6 Voice Donor languages ( † ) with 0% train–test speaker overlap; v2 for 13 re-split Makerere and Clear Global languages ( ⋆ )) before ( Zero-Shot ) and after ( Fine-Tuned ) monolingual fine-tuning. Fine-tuned entries report Mean±95% CI across N≥3 independent training runs per language ( N∈[3,9] ). Bold values indicate the better result between Zero-Shot and Fine-Tuned within each model pair.
Gemma-3n-E2B-it
Whisper-large-v3
MMS-1B-all
Language
ISO
ZS
Fine-Tuned ( ± 95% CI)
N
ZS
Fine-Tuned ( ± 95% CI)
N
ZS
Fine-Tuned ( ± 95% CI)
N
Acholi ⋆
ach
0.4656
0.2138 ± 0.0078
3
0.6487
0.1783 ± 0.0118
3
0.1848
0.1337 ± 0.0160
6
Akan †
aka
0.4380
0.1422 ± 0.0364
3
0.4543
0.1185 ± 0.0058
3
0.1914
0.1222 ± 0.0019
3
Amharic †
amh
0.5973
0.2113 ± 0.1159
6
1.5126
0.7286 ± 0.0077
3
0.2003
0.1614 ± 0.0146
3
Dagbani ⋆
dag
0.5082
0.1677 ± 0.0143
3
0.7623
0.1576 ± 0.0237
3
—
0.1555 ± 0.0008
3
Dagaare ⋆
dga
0.5017
0.1988 ± 0.0177
3
0.5478
0.1567 ± 0.0032
3
0.1915
0.1495 ± 0.0019
3
Appendix
Table 5: Character Error Rate (CER, decimal; lower is better) and number of independent training runs ( N ) per language on the speaker-disjoint WAXAL test sets before ( Zero-Shot ) and after ( Fine-Tuned ) monolingual fine-tuning. Fine-tuned entries report Mean±95% CI over N runs.
Model
ZS WER ( v2 )
FT WER ( v1 )
FT WER ( v2 )
Δ ( v2 − v1 )
Gemma-3n-E2B-it
0.9435
0.4150
0.4467
+0.0317 ( +3.17 pp)
Whisper-large-v3
1.1394
0.4612
0.4812
+0.0200 ( +2.00 pp)
MMS-1B-all
0.5146
0.3986
0.4199
+0.0213 ( +2.13 pp)
Appendix
Table 6: Speaker-Leakage Ablation ( v1 Speaker-Overlapping vs. v2 Speaker-Disjoint Splits across the 13 re-split Makerere and Clear Global languages). Over 94% of the WER reduction over Zero-Shot is preserved after removing all speaker overlap.
Gemma-3n-E2B-it (WER / CER)
Whisper-large-v3 (WER / CER)
MMS-1B-all (WER / CER)
Benchmark
Language
Zero-Shot
Fine-Tuned
Zero-Shot
Fine-Tuned
Zero-Shot
Fine-Tuned
FLEURS v2
Luganda ( lug )
0.8810 / 0.2304
0.6427 / 0.2039
1.0349 / 0.2795
0.5432 / 0.1448
0.3848 / 0.0756
0.4594 / 0.1107
Lingala ( lin )
0.6222 / 0.2154
0.5833 / 0.3326
0.8306 / 0.3036
0.2771 / 0.0910
0.1471 / 0.0420
0.1702 / 0.0460
Oromo ( orm )
0.9192 / 0.3186
0.7311 / 0.2712
1.0118 / 0.4449
0.6252 / 0.1773
0.6437 / 0.1724
0.7008 / 0.1877
Amharic ( amh )
0.8515 / 0.7144
0.4252 / 0.1768
1.3832 / 1.6091
1.1568 / 1.0720
0.2957 / 0.0824
0.3819 / 0.1057
Fula ( ful )
0.8710 / 0.3475
0.7165 / 0.3458
0.9009 / 0.3240
0.5419 / 0.1851
0.4891 / 0.1380
0.5007 / 0.1490
Appendix
Table 7: Out-of-Distribution (OOD) Generalization Word Error Rate (WER) and Character Error Rate (CER) across three independent benchmarks: FLEURS v2 (6 languages), Mozilla Common Voice ( v1 + lg , 4 languages), and Sunbird SALT (3 East African languages). Bold indicates better WER between Zero-Shot ( ZS ) and WAXAL -Fine-Tuned ( FT ).
Evaluated on WAXAL Test
Evaluated on FLEURS Test
Language
MMSWAXAL∣WAXAL
MMSFLEURS∣WAXAL
Diff
MMSFLEURS∣FLEURS
MMSWAXAL∣FLEURS
Diff
Luganda ( lug )
0.19
0.35
−0.16
0.39
0.46
−0.07
Lingala ( lin )
0.33
0.29
+0.04
0.15
0.17
−0.02
Oromo ( orm )
0.31
0.33
−0.02
0.64
0.70
−0.06
Amharic ( amh )
0.48
0.57
−0.09
0.30
0.38
−0.08
Fula ( ful )
0.44
0.58
−0.14
0.49
0.50
−0.01
Appendix
Table 8: Bidirectional Cross-Dataset Evaluation and Generalization Gap ( MMS-1B Adapter WER across 6 overlapping languages). MMSA∣B denotes MMS-1B with adapters fine-tuned on dataset A and evaluated on dataset B ; Diff reports the in-domain minus cross-domain WER difference ( MMSB∣B−MMSA∣B ).
Language
Samples ( N )
Pearson r
MAE
Mean Bias ( LLM−Human )
Akan ( aka )
50
0.74
0.13
+0.01
Amharic ( amh )
49
0.68
0.14
+0.05
Total / Overall
99
0.71
0.14
+0.03
Appendix
Table 9: Human vs. LLM-as-judge (Gemini 3.1 Pro) agreement on the continuous [0.0,1.0] semantic meaning-preservation scale across randomly sampled test utterances rated independently by native speakers of Akan and Amharic.
Multilingual speech foundation models claim broad language coverage, but how well they handle spontaneous African speech, and how best to adapt them, remains unclear. We benchmark zero-shot and fine-tuned speech recognition on the WAXAL corpus of spontaneous speech in 19 African languages, covering Whisper, MMS, Omnilingual ASR, and SeamlessM4T-v2. Fine-tuning lowers macro word error rate (WER) from 52.3 for the best zero-shot model covering all 19 languages (Omnilingual-7B) to 36.2 for a 300M-parameter MMS model and 31.5 for Omnilingual with an LLM decoder, although zero-shot models are already competitive on a few languages. Using Omnilingual's shared encoder, we compare CTC and LLM decoders with encoder pretraining held fixed. Between MMS and Whisper, the better model depends on the language family, but with a shared encoder the LLM decoder has lower WER on 18 of 19 languages, so the family pattern more likely comes from pretraining than from the decoder. The LLM decoder is, however, about five times larger and much slower on CPU. Pooling languages into one model, by family or across all 19, costs only 0.6 to 1.0 WER on average relative to per-language models, a modest price for serving all languages with one model. Fine-tuned models remain usable on read speech, and for held-out languages neither training on the other languages nor in-context examples replace in-language fine-tuning. A native-speaker audit across all 19 languages characterizes failure modes such as repetition loops. We release the code and a suite of over 100 fine-tuned models.
Victor Tolulope Olufemi, Oreoluwa Babatunde, Ramsey Njema +28
Automatic speech recognition (ASR) performs well for high-resource languages with abundant paired audio-transcript data, but its accuracy degrades sharply for most languages due to limited publicly available aligned data. To this end, we introduce WorldSpeech, a 24 kHz multilingual speech corpus comprising 65k hours of aligned audio-transcript data across 76 languages, collected from diverse public sources including parliamentary proceedings, international broadcasts, and public-domain audiobooks. For 37 languages, WorldSpeech provides more than 200 hours of aligned speech, with 28 exceeding 500 hours and 24 surpassing 1k hours. Fine-tuning existing ASR models on WorldSpeech results in an average relative Word-Error-Rate reduction of 63.5% across 11 typologically diverse languages.
Antonis Asonitis, Luca A. Lanzendörfer, Frédéric Berdoz +1
Recent large language models (LLMs) show strong speech recognition and translation capabilities for high-resource languages. However, African languages remain dramatically underrepresented in benchmarks, limiting their practical use in low-resource settings. While early benchmarks tested African languages and accents, they lacked exhaustive real-world noise and granular domain evaluations. We present AfriVox-v2, a comprehensive benchmark designed to test speech models under realistic African deployment conditions. AfriVox-v2 introduces "in the wild" unscripted audio for all supported languages. We also introduce strict domain verticalization, evaluating model accuracy across ten sectors including government, finance, health, and agriculture and conducting targeted tests on numbers and named entities. Finally, we benchmark a new generation of speech models, including Sahara-v2, Gemini 3 Flash, and the Omnilingual CTC models. Our results expose the true generalization gap of modern speech models in specialized, noisy African contexts and provide a reliable blueprint for developers building localized voice AI.