Research map / Large language models / Language, safety and interpretability

WAXAL: A Large-Scale Multilingual African Language Speech Corpus

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Speech and audio language models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces WAXAL, a openly licensed speech corpus of about 1,250 hours of ASR speech and 235 hours of single-speaker TTS recordings across 24 Sub-Saharan African languages.

Problem

Speech technology resources are scarce for most Sub-Saharan African languages, with few large-scale, high-quality, permissively licensed ASR and TTS corpora.

Approach

A multi-partner data collection of image-prompted natural speech for ASR and phonetically balanced studio-recorded scripts for single-speaker TTS, released under CC-BY-4.0 with speaker metadata.

Key results

• ASR corpus of about 1,250 hours of transcribed natural speech with 224,767 instances across 14 languages in the released data.
• TTS corpus of over 235 hours of single-speaker studio recordings across 13 languages (about 17,660 instances).
• Data released under CC-BY-4.0 on Hugging Face (google/WaxalNLP).

Surprising findings

• Only about 10% of collected ASR audio was transcribed, so most collected speech is unlabeled in the release.

Limitations

• Only about 10% of the collected ASR audio is transcribed.
• Dialectal and socio-linguistic variation may not be fully captured.
• ASR data is unscripted, so a small risk of offensive content remains despite manual QC.

Experimental setup

Datasets: WAXAL-ASR WAXAL-TTS

Closest papers by meaning

• Bridging the Gap: Enhancing LLM Performance for Low-Resource African Languages with New Benchmarks, Fine-Tuning, and Cultural Adjustments (AAAI 2025 · similarity 0.89)
• SwitchLingua: The First Large-Scale Multilingual and Multi-Ethnic Code-Switching Dataset (NeurIPS 2025 · similarity 0.88)
• WorldSpeech: A Multilingual Speech Corpus from Around the World (NeurIPS 2026 · similarity 0.88)
• BIGOS V2 Benchmark for Polish ASR: Curated Datasets and Tools for Reproducible Evaluation (NeurIPS 2024 · similarity 0.86)
• IndicVoices-R: Unlocking a Massive Multilingual Multi-speaker Speech Corpus for Scaling Indian TTS (NeurIPS 2024 · similarity 0.86)
• LAMA-UT: Language Agnostic Multilingual ASR Through Orthography Unification and Language-Specific Transliteration (AAAI 2025 · similarity 0.86)