Research · Text-to-Speech
This work presents the design and development of a neural Text-to-Speech system for Krio, a widely spoken language in Sierra Leone, as the first phase of a broader initiative to build speech technologies for African languages.
Abstract
Text-to-Speech technology enables computers to convert written text into natural-sounding speech. While high-resource languages such as English, Mandarin, and Spanish have benefited from significant advances in neural TTS, many African languages remain underserved due to limited speech datasets and linguistic resources. This research presents the design and development of a neural Text-to-Speech system for Krio and outlines a scalable path toward building speech technologies for additional African languages.
Recent advances in deep learning have enabled highly natural speech synthesis through neural Text-to-Speech systems. Models such as Tacotron, VITS, SpeechT5, XTTS, and Orpheus have demonstrated remarkable performance for well-resourced languages. However, most African languages lack sufficient datasets, pronunciation resources, and pretrained models, limiting their inclusion in modern AI systems.
Current TTS systems perform poorly for Krio because there are very few publicly available speech datasets, existing datasets are too small for training modern neural models, pronunciation varies across regions and speakers, there are no widely adopted pronunciation lexicons, and most pretrained models are optimized for high-resource languages.
The study aims to develop a high-quality Krio speech corpus, build a neural Text-to-Speech model capable of producing natural speech, design an automated preprocessing pipeline, evaluate speech quality using objective and subjective metrics, and create a scalable framework that can be extended to other African languages.
A speech corpus was developed using recordings from native Krio speakers spanning male and female voices, different age groups, multiple regional accents, conversational speech, narrative speech, number reading, general sentences, and everyday expressions. Each recording was paired with a verified transcript to support supervised training.
The preprocessing pipeline included audio normalization, noise reduction where appropriate, silence trimming, resampling to a consistent sampling rate, segmentation into training-ready clips, transcript validation, and removal of corrupted samples. This standardized the data and improved training stability and model performance.
The research explored modern neural TTS architectures such as SpeechT5, VITS, MMS-TTS, and Orpheus. The final system was selected based on synthesis quality, robustness, inference speed, and adaptability to Krio, with transfer learning from pretrained multilingual models helping reduce training time and improve performance with limited data.
Several obstacles shaped the work, including speaker diversity, emotional variability in expressive recordings, limited data compared with English corpora, and inconsistent pronunciation across speakers and regions. These challenges highlighted the need for careful data curation and efficient transfer learning strategies.
The model was evaluated using objective metrics such as Mel Cepstral Distortion, Word Error Rate, Signal-to-Noise Ratio, and inference latency, alongside subjective assessments of naturalness, pronunciation accuracy, intelligibility, fluency, and overall speech quality. The resulting TTS system supports voice assistants, educational tools, accessibility solutions, customer support, public information systems, audiobooks, language learning, and digital government services.
Future research will focus on multi-speaker TTS, emotion-controllable speech synthesis, voice cloning with appropriate safeguards, real-time streaming, speech-to-speech translation, cross-lingual voice transfer, and expansion to other African languages such as Mende, Temne, Limba, and Fula. This research demonstrates that high-quality neural Text-to-Speech systems can be developed for low-resource languages through careful dataset creation, preprocessing, transfer learning, and modern neural architectures.