The Intersection

Voice matching

Voice matching is the practice of generating AI content that sounds like a specific author or team, capturing structural patterns rather than surface tics.

Voice matching is what most people mean when they say make the AI sound like me. The naive version swaps a few stylistic markers. Use shorter sentences. Add the author's favorite adverb. Borrow their sign-off. The output reads like a caricature. Readers notice and it undermines trust.

Real voice matching works at a structural level. It captures how the author organizes arguments, which claims they hedge and which they make flat, where they break convention, and what they refuse to write. It pulls these patterns from a large enough sample of the author's existing work that the model can apply them without explicit instructions. The output is not a style transfer. It is a closer approximation of how the author would actually write a given piece.

The hard part is not the model. The hard part is the corpus. A voice-matching system trained on five blog posts will miss. A voice-matching system trained on several years of the author's writing, across formats and topics, has a real chance. The substrate determines the ceiling.

The Amdahl view

The best voice matching is the boring kind. The reader cannot tell the difference between the author's own writing and the AI-generated version. The worst kind is the flashy one that hits a few surface-level tics and misses the substance. Any voice-matching product that demos well on a single paragraph and falls apart on a longer piece is doing the flashy version. The test is always length. Short output is easy to fake. Long output reveals whether the system understands how the author actually thinks.

See customer intelligence running on your own customer conversations.