I test AI every day — the best AI voice generator I've ever used just dropped

Tom's Guide

2 days ago

You've probably heard an AI-generated voice by now. They are usually pretty noticeable, lacking in emotion or styling and still fairly robotic in nature. But with its latest update, Hume is changing that.
Evi 3, the latest version of the company's leading AI-voice generator, has just landed. While it has an array of pre-made voices to try out, the thing that really sets this model above the rest is its customization options.
Hume gave me access to an early version of the tool, allowing me to try out an array of custom voices and the ability to generate one down to the smallest details. Here's how it went.
Once you're on the Hume system, you get a few options. You can choose from a selection of pre-set voices, but that's not what we're here for. The new feature on offer is the ability to design a voice.
Do this and you'll enter a conversation with an AI. This will ask you questions about the voice you want to generate. Either you can take control, pitching an exact style of voice you want, or you can allow Hume to suggest options for you to take a guided approach.
I tried a variety of different options here, some pretty simple in nature, some overly complicated and specific. Impressively, most of the voices fit my descriptions pretty closely.
For example, the first voice I asked for was raspy and low in energy. It had a villainous nature, similar to a bad guy you would see in a fantasy film who rules over some evil kingdom.
Get instant access to breaking news, the hottest reviews, great deals and helpful tips.
Incredibly specific, but a few minutes later, I had a voice that fit this exact description chatting away to me. Not only is the tone matched, but also his vocal mannerisms, mocking me and being sarcastic in his tone.
Next, a British game show host, complete with an old-fashioned accent, a lot of energy, and an overly positive nature. Again, both accent and tone are matched up to the request surprisingly well.
I went on to try an array of different voices, ranging from pirate-like in nature to very simple American accents. All of which were achieved by Hume.
This isn't to say it was perfect throughout. Sometimes the voices would clip, revealing a slight robotic-sounding voice or a slip in the accent. It is also always clear you're talking to an AI, however, only slightly so, still sounding more human than robot.
You can also only design these specific voices by having a back-and-forth voice conversation with Hume AI. It would be great to be able to add text prompts to generate them, especially when you can't speak.
This also means it takes twice as long to generate a voice, having to work through a long conversation to get a result. It's a small but noticeable concern.
Compared to text, video, and image generation, AI voice generation hasn't seen the same push. Companies like Elevenlabs have been at the forefront, and the likes of Google and OpenAI have made progress in the field.
However, this is the best attempt I've seen in terms of customisation on the voices, not to mention in the tone and personality of said voices.
Hume claims that in a blind comparison against OpenAI's GPT-4o, EVI 3 was rated higher, on average, on empathy, expressiveness, naturalness, interruption quality, response speed, and audio quality.
The company also claims that the model outperformed GPT-4o, Gemini, and Sesame (a popular AI voice system) in ratings of how well it acted out a wide range of emotions and styles to study participants.
This, for now, puts Hume in a great place in the market. However, AI moves fast. While it currently stands out as a leader in AI voice generation, especially in terms of creative expression, they'll have to keep the updates coming to stay ahead.
You can try Hume's Evi 3 now on the Hume AI dashboard.