WAXAL: A Large-Scale Multilingual African Language Speech Corpus

arXiv:2602.02734 · eess.AS, cs.AI, cs.CL · Submitted 2026-02-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "WAXAL: A Large-Scale Multilingual African Language Speech Corpus".

Dev: Here's the gist of the paper. The WAXAL dataset introduces a large-scale, openly accessible speech resource for 24 Sub-Saharan African languages to address the digital divide in speech technology development The collection consists of an Automated Speech Recognition (ASR) dataset and a Text-to-Speech (TTS) dataset designed to catalyze research and enable the development of inclusive technologies.

Rosa: Dataset Scope and Components WAXAL is a largescale, openly accessible speech dataset for 24 languages representing over 100 million speakers The collection consists of two main components: an Automated Speech Recognition (ASR) dataset containing approximately 1,250 hours of transcribed, natural speech from a diverse range of speakers, and a Text-to-Speech (TTS) dataset with over 235 hours of high-quality, single-speaker recordings reading phonetically balanced scripts The ASR dataset spans 14 languages and the TTS dataset spans 13 languages Table 1 provides an overview of the statistics for both datasets.

Dev: Data Collection Methodology The WAXAL dataset was acquired through a multi-year effort (Jan 2021 - Mar 2024) funded by Google with involvement from several key partners in Africa, ensuring local expertise and community involvement The data collection involved distinct processes for ASR and TTS data 1.

Rosa: For the ASR dataset, participants were shown a diverse set of images covering at least 50 topics and asked to describe them in their native language to elicit more natural speech Each recording was captured in the speaker’s natural environment and has a minimum duration of 15 seconds A subset corresponding to 10% of the total collected audio was transcribed by paid, local linguistic experts using the local script where available or transliteration to English alphabet 2.

Title and authors: Rosa: We’re talking about the title "WAXAL: A Large-Scale Multilingual African Language Speech Corpus," and the authors include people from Google Research, University of Ghana, Digital Umuganda, Addis Ababa University, Makerere University, Loud and Clear Comm., Media Trust Ltd., Gates Foundation, AIMS Senegal.

Dev: That’s a lot of institutions backing this project. What matters is that they are pooling resources to create this resource for twenty-four languages representing over one hundred million speakers across the continent.

Rosa: The implication here is that they are trying to address the digital divide in speech technology development by creating an openly accessible resource under a CC-BY-four point zero license, which means researchers can actually use it widely <ref:2602.02734#pg1>.

Taro: I think making it openly available is important, but I always have to ask about the quality control on this scale of data; how much manual effort went into getting these one thousand two hundred fifty hours transcribed for ASR?

Dev: Well, the paper mentions that a subset corresponding to ten percent of the total collected audio was transcribed by paid, local linguistic experts using the local script where available or transliteration to English alphabet <ref:2602.02734#pg1>.

Rosa: That means they’re not claiming perfect transcription across all hours, which is important for setting expectations about what the models can actually learn from this specific corpus.

The paper's summary: Dev: Moving on to the summary, WAXAL is described as a large-scale speech dataset consisting of an ASR dataset and a TTS dataset designed to catalyze research and enable the development of inclusive technologies.

Rosa: So, essentially it’s about providing both the transcribed audio for speech recognition and high-quality recordings for text synthesis across twenty-four languages.

Taro: When you look at the scale, they have ASR with approximately one thousand two hundred fifty hours and TTS with over one hundred eighty hours of studio-quality recordings across those ten target languages.

Dev: The key part is the metadata they provide along with the data; this includes valuable speaker demographics, which helps researchers understand the diversity present in each dataset.

Rosa: They also mention that they've done a lot of work on ensuring linguistic quality by having local experts transcribe the audio using local scripts when possible.

Taro: But I’m still focused on what this means for the world; it means giving people in these languages a chance to have voice-enabled applications like virtual assistants or automated transcription services built for their specific tongues.

The paper's improvements: Rosa: The authors highlight several key aspects of the WAXAL collection as improvements, specifically mentioning the high-quality TTS corpus and the rich metadata they include.

Dev: They also point out that they’ve created a phonetically balanced script of approximately one hundred eight thousand five hundred words for each of those ten target languages to make sure the TTS data is consistent.

Taro: What I find compelling about the improvements is how they are explicitly designed to be used for building high-fidelity TTS systems, which is a specific need in these language communities.

Rosa: And they emphasize that the entire WAXAL collection is made publicly available under a CC-BY-four point zero license, which encourages widespread use in both academic and commercial research <ref:2602.02734#pg1>.

Dev: That licensing choice is significant because it lowers the barrier for entry for anyone looking to develop speech technology solutions for these languages.

Conclusion: Rosa: To wrap things up, WAXAL provides a foundational resource by offering around one thousand five hundred hours of annotated ASR and TTS data across twenty-four languages, which is a significant step toward addressing the scarcity that has hindered speech technology development for Sub-Saharan African languages.

Dev: It’s a vital resource for building and evaluating models because it gives researchers something concrete to work with when they are trying to develop robust systems for these specific linguistic characteristics.

Taro: From my side, I think what matters is that this collection allows people to start experimenting with creating voice-enabled applications in their native languages right now, rather than waiting for much larger datasets.

Rosa: So WAXAL is being released publicly under the CC-BY-four point zero license at Hugging Face so that the research community can access it and use it freely for building and evaluating models <ref:2602.02734#pg1>.

Dev: It’s a big contribution to making sure these languages get the speech technology they need to be inclusive, but we still have to be mindful of those limitations like dialectal representation.

Abdoulaye Diack, Perry Nelson, Kwaku Agbesi, Angela Nakalembe, MohamedElfatih MohamedKhair, Vusumuzi Dube, Tavonga Siyavora, Subhashini Venugopalan, 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

Google Research · University of Ghana · Digital Umuganda, Addis Ababa University, Makerere University, Loud and Clear Comm. Ltd., Media Trust Ltd. · Gates Foundation · AIMS Senegal

eess.AS, cs.AI, cs.CL

Submitted: 2026-02-02

Updated: 2026-11-14

Comments: Accepted at NeurIPS 2026 Evaluations and Dataset Track. Camera ready version

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 93/100

The gist: The WAXAL dataset introduces a large-scale, openly accessible speech resource for 24 Sub-Saharan African languages to address the digital divide in speech technology development The collection…

Key concepts

WAXAL Dataset
A massive, openly accessible collection of speech data covering 24 African languages. It includes transcribed audio for ASR and high-quality recordings for TTS, intended to support the development of speech technology in under-resourced regions.
ASR Dataset
This component contains about 1,250 hours of natural, transcribed speech from 14 languages. It is used to train systems that can automatically convert spoken words into written text, helping to develop robust speech recognition for diverse African linguistic features.
TTS Dataset
This collection consists of over 235 hours of high-fidelity recordings from community members reading balanced scripts in 13 languages. It is specifically designed to create synthetic voices that sound natural and accurate for Text-to-Speech applications.

Terminology

Summary

The WAXAL dataset introduces a large-scale, openly accessible speech resource for 24 Sub-Saharan African languages to address the digital divide in speech technology development The collection consists of an Automated Speech Recognition (ASR) dataset and a Text-to-Speech (TTS) dataset designed to catalyze research and enable the development of inclusive technologies. p. 0

Where the paper says it

To address this gap, we introduce WAXAL, a largescale, openly accessible speech dataset for 24 languages representing over 100 million speakers. The collection consists of two main components: an Automated Speech Recognition (ASR) dataset containing approximately 1,250 hours of transcribed, natural speech from a diverse range of speakers, and a Text-to-Speech (TTS) dataset with over 235 hours of high-quality, single-speaker recordings reading phonetically balanced scripts.

Dataset Scope and Components

WAXAL is a largescale, openly accessible speech dataset for 24 languages representing over 100 million speakers The collection consists of two main components: an Automated Speech Recognition (ASR) dataset containing approximately 1,250 hours of transcribed, natural speech from a diverse range of speakers, and a Text-to-Speech (TTS) dataset with over 235 hours of high-quality, single-speaker recordings reading phonetically balanced scripts The ASR dataset spans 14 languages and the TTS dataset spans 13 languages Table 1 provides an overview of the statistics for both datasets.

Data Collection Methodology

The WAXAL dataset was acquired through a multi-year effort (Jan 2021 - Mar 2024) funded by Google with involvement from several key partners in Africa, ensuring local expertise and community involvement The data collection involved distinct processes for ASR and TTS data

  1. For the ASR dataset, participants were shown a diverse set of images covering at least 50 topics and asked to describe them in their native language to elicit more natural speech Each recording was captured in the speaker’s natural environment and has a minimum duration of 15 seconds A subset corresponding to 10% of the total collected audio was transcribed by paid, local linguistic experts using the local script where available or transliteration to English alphabet

  2. For the TTS dataset, a phonetically balanced script of approximately 108,500 words was created for each of the 10 target languages Seventy-two male and female community participants were contracted for voice acting to ensure gender balance All recordings were conducted in a professional studio-like environment to ensure high audio fidelity and minimal background noise

Data Statistics and Scale

The total size of the released, ASR data is 1.7 TB, while the TTS data is 99 GB Table 1 provides an overview of the statistics for both datasets The WAXAL-ASR dataset has approximately 1,250 hours and spans 14 languages The WAXAL-TTS dataset has approximately 235 hours and spans 13 languages

Limitations and Ethical Considerations

The release of any large-scale human data requires careful consideration of its limitations and ethical implications

** Transcription Coverage:**

The current ASR release contains transcriptions for only 10% of the total audio collected by our partners

** Dialectal Representation:**

While efforts were made to recruit diverse speakers, the dataset may not capture the full dialectal and socio-linguistic variation present within each language

** Unintended Content:**

As the ASR data is unscripted, there is a small possibility of offensive or inappropriate speech, though this risk is mitigated by our manual quality control process

** Use-Case Specificity:**

The ASR dataset, with its diverse pool of speakers and environmental conditions, is not wellsuited for training high-quality, single-speaker TTS models

Conclusion

The WAXAL dataset represents a significant step toward addressing the resource scarcity that has hindered the development of speech technologies for Sub-Saharan African languages By providing around 1,500 hours of high-quality, annotated ASR and TTS data across 24 languages, we offer a foundational resource for building and evaluating models The WAXAL datasets are being released publicly to the research community under the permissive CC-BY-4.0 license at https://huggingface. p. 3

Where the paper says it

The WAXAL datasets are being released publicly to the research community under the permissive CC-BY-4. 0 license at https://huggingface.

co/datasets/google/WaxalNLP. p. 0
Where the paper says it

The WAXAL datasets are released at https://huggingface. co/datasets/google/WaxalNLP under the permissive CC-BY-4.

References

[1] Rosana Ardila, Megan Branson, Kelly Davis, Michael Henretty, Michael Kohler, Josh Meyer, Reuben Morais, Lindsay Saunders, Francis M Tyers, and Gregor Weber,“Common voice: A massively-multilingual speech corpus,” arXiv preprint arXiv:1912.06670,” 2019

[2] Alexis Conneau, Min Ma, Simran Khanuja, Yu Zhang, Vera Axelrod, Siddharth Dalmia, Jason Riesa, Clara Rivera, and Ankur Bapna,“Fleurs: Few-shot learning evaluation of universal representations of speech,” in 2022 IEEE Spoken Language Technology Workshop (SLT). IEEE,” 2023

[3] Mark JF Gales, Kate M Knill, Anton Ragni, and Shakti P Rath,“Speech recognition and keyword spotting for low-resource languages: Babel project research at cued,” in Fourth International workshop on spoken language technologies for under-resourced languages (SLTU-2014). International Speech Communication Association (ISCA), 2014

[4] Alan W Black,“Cmu wilderness multilingual speech dataset,” in ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE,” 2019

[5] Jorgen Valk and Tanel Alum¨ ae,“Voxlingua107: a dataset for spoken language recognition,” in 2021 IEEE Spoken Language Technology Workshop (SLT). IEEE,” 2021

[6] Xinjian Li, Shinnosuke Takamichi, Takaaki Saeki, William Chen, Sayaka Shiota, and Shinji Watanabe,“Yodas: Youtube-oriented dataset for audio and speech,” in 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE,” 2023

[7] Yifan Yang, Zheshu Song, Jianheng Zhuo, Mingyu Cui, Jinpeng Li, Bo Yang, Yexing Du, Ziyang Ma, Xunying Liu, Ziyuan Wang, et al., “Gigaspeech 2: An evolving, large-scale and multi-domain asr corpus for lowresource languages with automated crawling and transcription and refinement,” arXiv preprint arXiv:2406. p. 4

Where the paper says it

[7] Yifan Yang, Zheshu Song, Jianheng Zhuo, Mingyu Cui, Jinpeng Li, Bo Yang, Yexing Du, Ziyang Ma, Xunying Liu, Ziyuan Wang, et al. , “Gigaspeech 2: An evolving, large-scale and multi-domain asr corpus for lowresource languages with automated crawling, transcription and refinement,” arXiv preprint arXiv:2406.

11546,” 2024

[8] Moussa Doumbouya, Lisa Einstein, and Chris Piech,“Using radio archives for low-resource speech recognition: Towards an intelligent virtual assistant for illiterate users,” in Proceedings of the AAAI Conference on Artificial Intelligence, 2021, vol. 35

[9] Elodie Gauthier, Laurent Besacier, Sylvie Voisin, Michael Melese, and Uriel Pascal Elingui,“Collecting resources in sub-saharan african languages for automatic speech recognition: a case study of wolof,” in 10th Language Resources and Evaluation Conference (LREC 2016), 2016

[10] Daniel van Niekerk, Charl van Heerden, Marelie Davel, Neil Kleynhans, Oddur Kjartansson, Martin Jansche, and Linne Ha,“Rapid development of TTS corpora for four South African languages,” in Proc. Interspeech 2017, Stockholm, Sweden

[11] Josh Meyer, David Adelani, Edresson Casanova, Alp Oktem, Daniel Whitenack, Julian Weber, Salomon Kabongo Kabenamualu, Elizabeth Salesky, Iroro Orife, Colin Leong,, Perez Ogayo,, Chris Chinenye Emezue,, Jonathan Mukiibi,, Salomey Osei,, Apelete Agbolo,, Victor Akinode,, Bernard Opoku,, Olanrewaju Samuel,, Jesujoba Alabi, and Shamsuddeen Hassan Muhammad,“Bibletts: a large, high-fidelity, multilingual, and uniquely african speech corpus,” in Interspeech. 2022

[12] VAANI Team,“Vaani: Capturing the language landscape for an inclusive digital india (phase 1),” https://vaani.iisc.ac.in/, 2025

Appendix

Fig. 1 shows examples of some of the images used as prompts to elicit natural speech for the ASR dataset

Fig. 2 shows WAXAL: Distribution of the Transcribed ASR and TTS datasets in hours sliced by language

Fig. 3 shows WAXAL-ASR: Total utterances and hours of recordings per language

Fig.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Large-Scale ASR Model Training: The WAXAL-ASR corpus provides 14 languages with approximately 1,250 hours of transcribed, natural speech suitable for training and evaluating ASR models. This enables the development of robust Automatic Speech Recognition systems capable of handling diverse African linguistic characteristics like tonal distinctions and complex morphology.

  2. High-Fidelity TTS System Development: The WAXAL-TTS corpus offers over 180 hours of studio-quality recordings from members of the community reading phonetically balanced scripts for 10 languages. This data is specifically designed for building high-fidelity Text-to-Speech systems, allowing for the creation of synthetic voices that sound natural and accurate in those languages.

  3. Inclusive Technology Deployment: The collection serves as a vital resource for the digital preservation of these languages and can be used to catalyze research, enable the development of inclusive technologies. This facilitates the creation of voice-enabled applications, such as virtual assistants or automated transcription services, in Sub-Saharan African languages.

Abstract

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.

Sources

Related papers