A research team at The University of Texas at El Paso (UTEP) has identified a feature of speech that could help artificial intelligence (AI) voices sound more natural. The study, led by Nigel Ward, Ph.D., professor in UTEP’s Department of Computer Science, found that people tend to articulate words less precisely when expressing positive feelings. This phenomenon, known as phonetic reduction, has not previously been considered significant by speech technology researchers.
“People often perceive AI systems as cold and unaware and don’t trust them, even if their task performance is superb,” Ward said. “Current AI voices are consistently highly intelligible, so there’s space to sacrifice a little articulatory precision when needed. We plan to build voices that do this, so that AI systems may be finally able to escape the cold, robotic stereotype and become more communicative, more trustable, and more useful.”
The findings are detailed in the peer-reviewed paper “Phonetic Reduction is Associated with Positive Assessment and other Pragmatic Functions,” published in Speech Communication. Undergraduate research assistants Raul O. Gomez, Carlos A. Ortega, and Georgina Bugarini coauthored the paper.
Researchers recorded speakers saying one of six phrases twice—once neutrally and once positively—and had non-expert judges compare each phoneme between the two versions. In English samples, speakers were about a third more likely to reduce pronunciation in positive tones compared to neutral ones; 30% were judged “reduced” and 9% “highly reduced.” Similar results appeared in Spanish: 35% reduced and 4% highly reduced.
Phonetic reduction also appears in speech serving other social purposes such as self-talk or expressing uncertainty. However, this study focused on its correlation with positive speech.
“When you’re speaking in a more positive way, there is a higher pitch and you tend to speak a little bit faster,” said co-author Raul Gomez. “We made sure to eliminate that by making speakers do the positive one first and then the neutral reenactment, so in case the neutral one was longer, they could shorten it down. So we’re only viewing the reduction via the positiveness in the voice rather than the length.”
Ward’s lab uses ReduEst—a tool developed by master’s student Javier Vazquez—to estimate levels of phonetic reduction in speech data. The software is available for free for other researchers interested in similar studies.
“We want to move away from just a transactional translation system, where we just kind of get information across, and more so into a conversational system, where we can express feelings and emotions,” Vazquez said.
Ward noted realistic speech could be particularly helpful for automated voice systems handling sensitive situations or embodied AI agents like robots or self-driving cars operating in dynamic environments.
In September 2026, Ward’s team will attend Interspeech 2026 in Sydney to encourage further research on phonetic reduction across languages beyond American English and Mexican Spanish.
“Even though we might not be able to conduct research on those languages, other people can pick up the torch and then maybe corroborate the findings – or find something else,” Ward said.
