Dissecting Speech, Manufacturing Difference

Artificial intelligence (AI) technologies that process the sounds of human voices—machine listening and text-to-speech (TTS) systems—are increasingly being implemented across high-stakes infrastructures spanning government, healthcare, security, and media. Central to these systems’ operations is the reconstruction of the human voice to fit machine learnable formats. In this chapter, Kang and Becerra Sandoval examine one such process of reconstruction in the development of AI voice generation systems: the partitioning of human speech across what they describe as the “acoustic-linguistic divide”. Specifically, the chapter analyzes how this imagined boundary between the acoustic and linguistic components of speech falters in the development, training, and deployment of two distinct but related AI voice generation applications: AI-enabled accent modification and “diverse” synthetic voice generation. Ultimately, the authors argue that across these two contexts, the acoustic-linguistic divide works to reaffirm and embed existing social hierarchies of difference in their underlying technical architectures.