A twin sounds like you for two reasons. First, it answers from your own material, which is full of your phrasing, your examples and your way of reasoning, and a language model is very good at matching a style it can see in context. Second, during setup you answer guided questions about your voice (how formal you are, how you open and close, what you would never say) and those answers are given to the model with every conversation. Voice is not a cloned recording; it is your point of view, carried faithfully.
What "voice" means to a language model
Language models learned from an enormous range of writing, so they are unusually good at noticing and reproducing a style: sentence length, warmth, directness, the kind of examples a person reaches for, how they handle disagreement. Show a model a few thousand words of someone's writing and ask it to answer as that person would, and it will match the manner closely. It is not imitating a recording; it is continuing a pattern.
The two carriers of your voice
1. Your material
Every answer your twin gives is assembled from passages of your own writing or speaking. The passages bring your phrasing with them. A pastor who always starts with a story will have a twin that starts with a story, because the sermons do. A coach who is blunt will have a blunt twin. This is why material in your own words matters so much more than summaries about you.
2. The guided voice questions
During setup, AccessMe asks a few questions about how you speak: how formal you are, how you open and close a conversation, what you would never say. Your answers are placed in front of the model with every conversation, so even when the retrieved material is a dry lecture note, the twin frames it the way you would in person. You can revisit these answers at any time.
Point of view is the harder part
Sounding like you is mostly style. Thinking like you is point of view: what you emphasize, what you dismiss, where you refuse to give a simple answer. This comes almost entirely from material, and especially from your own written answers to the questions you are asked most. A twin given only your books will know what you wrote; a twin given your answers to real questions will know how you apply it. See what to upload.
What voice does not include
A twin does not use your actual voice. There is no voice cloning, no audio and no phone channel; a twin is text chat. "Voice" here means the written manner and perspective, which is where trust actually lives.
Tuning it
- Too generic? Add more material in your own words, especially conversational answers. Generic answers usually mean the twin had little of yours to work with on that topic.
- Too formal or too casual? Revisit the voice questions.
- Right words, wrong emphasis? Write out how you would actually answer that question and add it. The twin will follow your lead next time.
Across languages, the voice carries too: someone can talk with your twin in Spanish, Korean or Portuguese and still get your manner and your point of view, even if your material is in English.
Common questions
Does a twin clone my voice?
No. There is no audio or voice cloning; a twin is text chat. 'Voice' means your written manner, tone and point of view, which the twin carries from your material and your answers to the guided voice questions.
Will my twin sound like me in other languages?
Yes. People can talk with your twin in their own language and the twin keeps your manner and your point of view, even when your material is in English.
My twin sounds generic. What should I change?
Add more material in your own words on the topics where it sounded generic, especially conversational answers to real questions, and revisit the voice questions. Generic answers almost always mean thin material on that topic.