The End of Accents (www.wired.com)

AI-driven “accent training” and real-time neutralization tools—exemplified by startups like BoldVoice (whose Accent Oracle correctly guessed the author’s Korean origin) and vendors such as Krisp and Sanas—are now automating what used to be speech coaches’ work. The author recounts using BoldVoice’s diagnostic, which produced scores (e.g., 89–92% “Lightly Accented” to “Native or Near-native”) and flagged concrete phonetic targets: the English th sound, final-consonant devoicing (making did sound like dit), and vowel length distinctions (seat vs. sit). The piece also summarizes why accents persist from a technical perspective: different languages have different phoneme inventories (English ≈44 phonemes, Korean ≈40), so learners’ brains substitute nearest sounds, producing recognizably “foreign” patterns. For the AI/ML community this raises practical and ethical questions. These models can boost workers’ employability—call centers already use accent-smoothing to tailor voices to customers—but they also enable “digital whitewashing,” erasing vocal identities and amplifying existing accent hierarchies. A 2022 UK study cited finds a quarter of workers experience accent discrimination and nearly half have been mocked. The story highlights trade-offs for designers: improve accessibility and economic outcomes without reinforcing linguistic prestige biases, and ensure transparency, consent, and cultural sensitivity when building models that change how people sound.