Unhurried

I stammer. When I talk to a voice bot, I can get stuck on a word for a few seconds. To me, that is the middle of my sentence. To the bot, that silence means I am done, so it answers a question I never finished. Then I have to start again, and often it happens again. Unhurried is a layer that sits between a person and any voice AI. It does three things: 1. It works out when you have really finished. A silent pause alone does not end your turn. Unhurried looks at your words, then at your voice, such as pitch and how your sound fades. If it still isn't sure, Muse Spark makes the call, with a limit of 700 ms so the conversation doesn't drag. While it waits, the screen says "take your time". It also learns how long your own pauses usually are. 2. It removes repeated sounds without changing what you said. "K-K-Kamal" reaches the bot as "Kamal". Code checks every result, so the model can't add words you never said. If it tries, the change is blocked. 3. It checks before anything important. Before it books, pays, cancels or sends, it asks you to confirm. If it got something wrong, you can correct it first. I also built Voice Studio for recordings. You upload your voice and it finds the repeated sounds and cuts them out, so the result is still your own voice. Muse Spark listens twice, Meta's wav2vec 2.0 finds the exact timing, and Muse Voice Transcribe gives a third check. You can see each cut, and your choice always wins. There is also an optional AI copy of your voice, which is watermarked and locked with a passphrase. Does it work? I tested it on 15 recordings of my own free speech, 8.8 minutes in total, streamed through the live server. A normal voice bot replied while I was still talking 14 times, in 8 of the 15 recordings. With Unhurried, that dropped to 2 times, in 2 of the 15. Invented words: 0. This has limits. It is one person's voice and a small set, and Unhurried still cut me off twice. The speech service also still reports the end of speech too early about as often as before. Unhurried doesn't fix that signal, it just stops trusting it blindly. The results page shows all of this. What it is not: it does not try to fix how anyone talks. It never scores your speech and never gives speaking tips. The only thing I measure is the AI: how often it cuts in, whether it understood, how much delay it adds, and whether it invented anything. Your live audio is not stored unless you turn that on. Why it matters beyond stammering: people also pause when they are thinking, speaking a second language, or finding a word after a stroke. Voice AI is being built into more everyday services, and it should not only work for people who speak without stopping. If it works for a stammer, one of the hardest cases, the same layer can help many others. Built with: Muse Voice Transcribe, Muse Spark, wav2vec 2.0, Python, FastAPI and WebSockets, with plain HTML and JavaScript in the browser. All code was written Oct 28 to Nov 4, 2026.