Back in the day after school, one of my favorite things to do was to listen to the radio, with a fresh cassette ready to push the record button as soon as a song I liked would come up. Often that took quite a while, and in the meantime I would record random stuff and play it back to me. The sound of your own recorded voice takes some getting used to. And the human voice sounds different to you and to others when you talk (or sing, for that matter).
Yesterday a new model called Dessa was presented (there are plenty of samples on the link to listen to). It’s a voice model out of Jack Dorsey’s (the guy that founded Twitter) AI lab at Block, Inc., and it can run locally on any computer. It sounds and feels like a human voice. Including all the fillers, pauses, changes of voice, and empathy you would find in a real human conversation.
I remember trying out Eleven Labs three years ago and experimenting with different characters. It was really good back then and has gotten way better, but you still knew it was AI. There was a certain artificialness still lingering, lacking the rough and imperfect aspects that make human speech, well, human. The sudden pauses, the slight cough, the interruption, the euphoria, the humor, and the patterns that each of us have once we are on the phone. There are some people that are really into recording voice messages and sending those around (I’m not one of them, to each their own), and most voice messages are filled with those little sprinkles of human thought and interruption.
Giving that humans have been speaking to each other for well over 100.000 years, the first recording of a human voice is pretty new. Starting 170 years ago with the phonautograph by Eduardo Leon Scott and with the phonograph, gramophone, and eventually the magnetic tape recorder that led to a quality of recorded voices that you could actually understand when played back.
Fast forward to the present: When thinking of ways computers might speak to us, there is a certain expectation. Friendly, intelligible, witty. But it is still clear that you are talking to and listening to a computer. A lot of voices have undoubtedly been inspired by this guy:
Hal 9000 in 2001 Space Odyssey
How does that all change if you have the genuine feeling you are on the phone with a real person?
Not a very good and clean service agent who is doing customer service or booking your trip. But someone that makes pauses, giggles at a choke, or gasps at exciting news. That is lacking the perfection that is sought out in efficiency and repeatability.
Like talking to your spouse, buddy, or coworker. Dessa was not trained on text, like the LLMs in use now, but on human speech. And it is able to speak and listen at the same time (which is something many humans have a hard time doing, frankly).
It is not available publicly and is still in research stage. But it adds a new layer to the possibilities and threads of all that is AI.