Tone-to-text applications for sub-Saharan African tone languages

Based on experimental findings, this paper proposes a framework for computer recognition of speech tones in Niger-Congo languages. This language family includes over 1000 languages in Sub-Saharan Africa and approximately one billion speakers. Many of these languages use pitch contrast to differentiate words in a system of two or more pitch levels (e.g., high, mid, and low). The results of two new studies conducted in Nigeria from 2013-14 indicate speech-tone is perceived as tonemic intervals (e.g., high-high, high-mid, and high-low). An experimental study (n = 1448) identified ranges of pitch difference that form perceptual categories for these tonemic intervals. An initial application of these findings is adding tone diacritics to text by interpreting fundamental frequency. Representing tone in text has been a persistent problem for Niger-Congo languages and smartphones are ill-equipped for marking tone diacritics. A signal processing application that recognizes speech tones and marks them on computer text would be more efficient than manually highlighting and marking each syllable. Frequency distribution of tones is dynamic between speakers. Thus, there are potential benefits of using machine learning to create speaker-dependent software. Because the proposed method relies on existing algorithms for fundamental frequency, the problem of estimation errors will also be addressed.