Synthetic Voice Data for Automatic Speech Recognition in African Languages

TL;DR

This study addresses the lack of automatic speech recognition (ASR) resources for Africa’s vast linguistic diversity by proposing a three-step pipeline: large-language-model (LLM) driven synthetic text, text-to-speech (TTS) synthesized speech, and ASR fine-tuning. It demonstrates that synthetic voice data can be produced at a fraction of real-data costs (often <1%) and can substantially improve ASR performance for several languages (notably Hausa, with mixed gains for Dholuo and Chichewa) when combined with real data in carefully chosen ratios. The work also highlights critical challenges in synthetic-data quality, including inter-coder reliability in language evaluation, downstream biases, and non-standardized scripts that complicate evaluation, and it emphasizes the need for robust evaluation protocols. All data and models are publicly released to accelerate follow-on research and practical deployment for African languages. The results suggest synthetic data is a promising complement to human data, capable of enabling broader access to speech technology across low-resource languages, while underscoring the importance of dataset quality and evaluation rigor.

Abstract

Speech technology remains out of reach for most of the over 2300 languages in Africa. We present the first systematic assessment of large-scale synthetic voice corpora for African ASR. We apply a three-step process: LLM-driven text creation, TTS voice synthesis, and ASR fine-tuning. Eight out of ten languages for which we create synthetic text achieved readability scores above 5 out of 7. We evaluated ASR improvement for three (Hausa, Dholuo, Chichewa) and created more than 2,500 hours of synthetic voice data at below 1% of the cost of real data. Fine-tuned Wav2Vec-BERT-2.0 models trained on 250h real and 250h synthetic Hausa matched a 500h real-data-only baseline, while 579h real and 450h to 993h synthetic data created the best performance. We also present gender-disaggregated ASR performance evaluation. For very low-resource languages, gains varied: Chichewa WER improved about 6.5% relative with a 1:2 real-to-synthetic ratio; a 1:1 ratio for Dholuo showed similar improvements on some evaluation data, but not on others. Investigating intercoder reliability, ASR errors and evaluation datasets revealed the need for more robust reviewer protocols and more accurate evaluation data. All data and models are publicly released to invite further work to improve synthetic data for African languages.

Figures (7)

• Figure 1: Overview of process to create synthetic voice data, including three key steps and human evaluation.
• Figure 2: Synthetic text generation prompt to generate simple sentences in target language and English translations
• Figure 3: Comparison of readability [1..7] scores for synthetic text generated by various LLMs for 10 African languages.
• Figure 4: Mean Kanuri readability and naturalness score by rater sample size. The shaded regions represent 95% confidence intervals derived from bootstrap analysis.
• Figure 5: Heatmap of mean Kanuri readability scores by individual linguists. Each cell displays the average score assigned by each linguist rater for sentences generated by different models.