Is there any good local voice that is equal to GPT Live?
I've dreamed for years of having an orchestrator like ChatGPT Live — running 24/7, always there, so I can just talk to it whenever I want without that Hey Siri bullshit. Right now I have a button on my Apple Watch: I press it, talk to my agent, and it works.…
I've dreamed for years of having an orchestrator like ChatGPT Live — running 24/7, always there, so I can just talk to it whenever I want without that Hey Siri bullshit. Right now I have a button on my Apple Watch: I press it, talk to my agent, and it works. But I want that nice full-duplex goodness fast responses, and a voice that can make tool calls and orchestrate my smarter agents. How far are we from a local ChatGPT Live? One that actually works and doesn't sound robotic and boring like every other one I've tried? One that can ignore other voices and jump in as i talk, read a reply from smarter agents back to me? I currently use Supertonic because it starts instantly and is the fastest thing out there. I need speed and quality at the same time, and Supertonic/pocket TTS only has the speed but the quality is flat and lifeless.
已收录讨论
I think the best bet for this is Moshi duplex model, otherwise personaplex by Nvidia. What's your setup? Also I would like to know what's your apple watch setup like OP, I'm also interested in having a local voice that can take actions etc.
Agree, this. But its not gonna be real time full-duplex.
I started with a simple native Apple shortcut where I make a post request to my server. Basically, I just record my voice, and Apple’s native transcription transcribes the voice, and that text gets injected directly into Claude code. You can have that in 10 minutes; just tell your agent. Then, I decided to make an Apple Watch app where I just want to have the app running with one big invisible button that I just press and talk forever, no limit. I also have a screenshot button so if I talk, it automatically sends a screenshot to my Claude code + text.
Not real time for you doesn't mean not real time for everyone
I tried Qwen3-TTS with streaming enabled on my mac m1, but all default female English voices sound so weird and Chinese. Is there a library of good cloned voices that can be tested with Qwen3-TTS?
I do 100% agree with you, the way it speaks English really hurts my ears 😱. As I have just switched to it, I'm still trying to figure out what and if anything can be done to improve it.
I have recently switched from Kokoro-82M to Qwen3-TTS due to its voice design capability and much better emotion and expression support, albeit a little slower than Kokoro, it's still practical enough to support real time conversations, just not full duplex. I do hope sooner there will more full duplex model available, right now, I only heard about Personaplex and Moshi, haven't tried any of them yet.
highest quality? = fish s2 pro
I think the missing feature is interruption support. GPT Live feels natural because you can stop it it stops talking away remembers where the conversation was and picks up smoothly. Most local assistants still feel like they are waiting for their turn of having a real conversation.
Ward!
Siri voices are quite good you can use them using the say command on mac
Vui and pipercat might be worth looking into.