I had preview access to this one for a few weeks. It's very good. I had one conversation that lasted a full hour while I was walking the dog, got some good brainstorming done against one of my projects.
The best feature is that it can delegate questions out to GPT-5.5 in the background, so you're no longer restricted to a voice model that's several years behind the frontier.
I did report a fun bug with it though: it was interrupting me and laughing at my (not really intended as) jokes while I was still talking! They seem to have clamped that behavior down thankfully, it felt a bit rude and condescending.
Is it responsive to personality settings? I actively don't want fake AI girlfriend, but I do get a ton of value out of voice mode. Looking forward to trying this but hoping it's not a creepy overdone mess (like Sesame). Expectations are they'll keep doubling down on fake AI girlfriend approach because the thing I want probably wouldn't drive engagement anywhere nearly as well
I have my Chat personality settings stripped right down to no-fluff. I'd want voice to be more akin to the Star Trek computer, and less akin to as you said an AI friend, but previously it was tuned too personable/friend-like.
Star Trek computer voice model is something I have yet to encounter, and I've looked repeatedly :) It's not about a specific voice, it's the fact they managed to capture "I am a utility" perfectly in the voice. Our modern friends do not want to be thought of as a utility, but to engender trust and agency all of their own and that's a huge problem for me.
Prompt text: There are 14 varieties of tomato soup available from this replicator. With rice, with vegetables, Bolian style, with pasta specify hot or chilled.
I think the request here is not about sounding like Majel Barrett but in keeping the output extremely terse and unobtrusive.
There's been a few studys showing that novices love LLM output that's long, but experts hate it. As an example, I've been tasked with using some agentic PM tool to write specs, and it keeps generating these huge page long outputs with "HBR voice" bolded summaries of paragraph long bulletpoints. I.e.:
> Right-size hard, and watch the one open-ended edge. Endorse the DRI's simplifications wholesale: drop the runbook-per-alert mandate (keep 1–2 diagnostic-only runbooks for the high-priority set), and ride durability on the existing weekly incident + monthly operational reviews — no new governance. The single scope-creep risk is the coverage strand (gaps are defined by absence); bound it to gaps evidenced by real, already-missed customer-facing outages, not a proactive gap hunt. Curing ownership gaps (e.g. foo-bar, no clear owner) is finite in-scope work.
There's dozens of these every iteration. I can't imagine trying to deal with that via voice, I would just zone out after the second sentence.
I'm just going to say, Wow. That's a pretty incredible result. I had no idea you could clone a voice with such a short snippet of audio these days.
I've had this dream of talking to the Enterprise-D computer since I was 8 years old. Midlife-crisis me still has that dream, but hooked up to Home Assistant so it can actually do useful things too. A couple months back, I went looking around for "clean" samples of Majel's voice as the computer but didn't have a lot of luck. Even though there are three television series and several movies, pretty much all of them have some amount of background noise, bleeps, bloops, or warp core thrum. (As this one does.) There may be modern ways to clean those up without affecting her voice much, but I haven't dug into that yet.
There are a few audiobooks narrated by Majel Barrett but obviously her role as the computer was proper voice acting and so the books would not be good source material.
There were also a few games/CD-ROM (Omnipedia) with some samples, but they did not bother to post-process them for that lofi Enterprise-D computer feel. Can _probably_ be replicated fairly faithfully after the fact, but I only know a _little_ about audio post-processing.
According to her son (Rod Roddenberry), Majel sat down in a studio and recorded audio samples specifically for the purpose of having her voice cloned someday for future Star Trek stories. However, those haven't been released publicly. (And likely never will, but I can dream, can't I?)
Edit: I played with your samples and the time it takes to generate the output audio is pretty brutal. Too slow for interactive use. Maybe that's a limitation of the HF-hosted app, though.
If you read their GitHub README and the code, it's possible to separate and cache the voice cloning step, and they have some streaming variant with low first chunk latency. I only played with it via Hugging Face. It is astonishingly good with certain rare accents, and it responds well to longer input clips
I noticed that too, I guess the content in the training data suggests that when some text self-identifies as "cold, hard logic" it additionally assumes a mean tone. Additionally it makes responses more likely to disagree even when given inputs that are more or less valid. A while ago I tested out giving some idea to a model with the "hard logic" prompt, then taking its own output, reverting the conversation and giving it its own statement again, making it disagree with the very same statement it just made itself. This was mostly just humorous, and doesn't indicate too much, after all both statements might've been wrong in some way, but the statement itself was moreso an opinion with no fully right answer, showing this tendency to disagree.
I guess when we tell models to use "cold logic" they don't interpret that purely definitionally, but moreso with what this statement usually implies, which is often a disagreement between two people, one trying to leverage supposed logic to disagree and denigrate the other party, oftentimes actually arguing out of emotion and not the logic they claim to use. This probably occurs enough to give model responses a mean tint. That's my theory on why this could occur at least.
One popular speech synth from back in the day, I believe it was WillowTalk, had a voice called Colossus, which sounded like the voice module of the computer from Colossus: The Forbin Project. This voice was used for that of CATS in the famous "All your base are belong to us" Flash video.
Another WillowTalk voice was a clone of DECtalk's Perfect Paul good enough to be used as the voice in the MC Hawking rap recordings.
It came from a company called Willow Pond Software. That seems to narrow searches down.
WillowTalk is apparently still of interest to Half-Life modders because another WillowTalk voice, possibly a clone of DECtalk's Huge Harry, was used as the Black Mesa VOX facility-wide announcement system in the original Half-Life.
I think that's mostly just a frequency shift :) You could probably recreate it with another model and some effects on top. Also, why the hell not for voice mode haha.
I explicitly prompt my small local models to channel that kind of energy. It's one of few ways to get them to just spit it out without yapping on about temperature- and humidity-appropriate activities when I just asked if it'll rain this week.
You do have to do it carefully and not literally say "Star Trek" or else it'll start yapping about main shields being at full strength.
It’s 2026 what the customer wants is probably an indication of what not to do if you’re a hyperscaler.
Whoever ends up actually winning after the crash will be the ones who figure out the needs of users, not the needs of the financial machinations of the companies trying to grab land.
> "The best feature is that it can delegate questions out to GPT-5.5 in the background, so you're no longer restricted to a voice model that's several years behind the frontier."
Wow. That's exactly what I hoped they would do.
This issue has held me back from using ChatGPT's voice mode as much as I otherwise would have, because I also use it for brainstorming while commuting, exercising, etc., and don't want it to feel stuck in the past.
Funnily enough - I built this (delegation) over the weekend with Fable for a local voice chat running 100% on local LLMs, Parakeet and Kokoro. I say "...ask the thinking model..." and that redirects it to Qwen 3.6 27B on vLLM.
Can't claim originality though - it was inspired by Sesame - where their models will invoke a search, or check the weather etc, and make a vocalisation to keep you engaged.
Turn taking is one of the hardest things to get right for the exact reasons mentioned - but does seem to be the way that Claude.ai's voice works - in a very obvious way.
Anthropic + OpenAI both rug-pulled voices I liked and got used to and OpenAI really dumbed down their voices at the same time - Arbor went from Estuary English and almost "jack the lad" to some generic English accent. Claude had a Birmingham accent and said things like "shit", ending sentences like "So you're telling me that they asked for a 90% discount yeah?" - then it changed overnight to a mock Derbyshire accent with a dull tone.
ChatGPT's voice also gaslights me for conventional opinions - "my Eastern European neighbour helped me lift a wardrobe upstairs - something you just can't ask your typical neighbour neighbour"... then you get a full on left-leaning lecture from the safety layers rather than a head nod or "what luck!"
Claude + Sesame are nowhere near as overbearing.
In both cases - from edgy and engaging to something that just didn't gel.
The point of making my own assistant is that I can talk for as long as I want, episodic memory is personal and private, there's no "trust me bro, we're a big corporation" vibes.
This was not my first attempt - when I had a bunch of Opus credit around Jan/Feb - I tried really hard and created something that was not good enough. What I have now, is working, and each session is training Claude/Codex on what to tune, and to fix.
"Just had a convo - can you look into what happened?" And if it's one I don't mind sharing with the model - I'll say, "and what did you think of the questions I asked?" Sometimes it'll give a lovely commentary on how the model did.
af_heart is probably the smoothest voice - but yes it's more like another commented - more "StarTrek" than "telesales assistant that pauses and laughs at your jokes".
If you're on a similar path and want something full duplex - the go to solution is PersonaPlex from Nvidia based upon Moshi.
Seems a bit knee jerk. I go on walks and bike rides all the time. A couple of times I’ve used voice mode and it’s been interesting. I could have listened to music or a podcast, listen to YouTube or just unplug. But every walk is different.
If you’re uncomfortable with this new world, and I’m sure I am even as I participate, you could tell us more about that?
This may not be you, but the people who have said that to me irl walk significantly less than I do. And saying these things to people who don't yet exercise can make them far less likely to start as it's a far bigger step to do so. Podcasts and recorded lectures are what got me exercising in the first place, because I was excited to hear the next part. I now only have headprones on some of my walks, but the gateway drug of entertainment was a very useful stepping stone. And still often better than sitting at a desk.
It's coming from a place of objecting to burnout/overwork culture.
Recently I witnessed a CTO mention in a public channel that with Claude remote control people can now work while getting a coffee or other breaks during the day.
Tech is actively moving in a direction of destroying all the gains from the labour movement in service of enriching capital out of a combination of FOMO or fear of being replaced.
So yeah, when folks say "hey look now I can even work during leisure activities!" yeah, the reaction is negative.
Workers in 2026, even non-tech workers have an easier life than workers alive at any other point in human history.
Yes ofc there are problems - collectivist land use laws that ban construction near homeowners continue to drive housing costs higher for example. But if you asked any worker today if they would have preferred to be alive 40 or 400 years ago I would be shocked if any said yes.
So I'm clear: as long as our jobs don't devolve to 100 hours a week in a sweat shop or tilling the earth for the lord of our fiefdom, we should just sit back and be happy?
Is that what I should take away from your comment?
>It's coming from a place of objecting to burnout/overwork culture.
Sure, but that's not everyone. I want this because it's hard to carve out time for my fun side projects. :)
Also, not everyone feels burnt out or overworked, and may find their actual work enjoyable. I've definitely had many showers where I've been thinking about interesting work problems and never felt particularly burdened by it (quite the opposite, in fact).
Sure but context matters and given this is Hacker News, a lot of discussion centers around tech as a profession (and that's doubly true for AI adoption). You can forgive folks for jumping to natural conclusions. :D
It reminds me of that old parable where the acolyte asks the priest "Father, can I smoke while I pray?" and gets told "No, you should be focusing on praying. That wouldn't be respectful". But then he says "Can I pray while I smoke?" and gets told "of course, you can pray at any time".
If it's 'walking time', I probably don't want to consume that with work. But if it's 'working time', it could be great to have a nice walk during it.
Sometimes I want to learn something while I walk. Sometimes I want to listen to stories. Sometimes I want to just be present. All are good. I highly recommend letting people enjoy whatever they enjoy.
Given the personality type common on HN, I imagine that the GP, even if unplugged from all technology on their walk, wouldn't be in a mindful state of enjoying their surroundings, but rather would be "lost in the clouds", stewing on the same ideas/thoughts/problems; but with those thoughts going more in circles, due to a lack of ability to verify anything.
Literally me. Everyone is different, and that's fine. But I don't have the privilege of living in an area where I can talk to people about the things I am thinking through. It's very rural. Having a _utility_ that can act as a sounding board while I spew out my thoughts on a walk is a really meaningful improvement to my current situation.
there's no benefit to anyone for being so reductive, from whichever side you're on.
As a progressive person, the constructive thing is to try to see how to harness it for good or how to fight it, it's not to pretend that it's weak. It's not weak.
Not saying you aren't good at your job but this statement gives me pause. Given what I see on the daily as output from models I hope there's still a lot of human interrupt behind the wheel with respect to the design of drugs.
Every drug indeed has a huge burden of proof before it is distributed to humans, including the 3 levels of FDA clinical drug trials that cost billions of dollars, as well as all the internal testing we do beforehand.
Is it a dumb-down version of GPT like the current voice model? At least in french, I find the current GPT voice mode to be useless, to the point I only use the dication mode. I would ask a question and it would answer something along "That's a interesting question. I can help you with that. Anything you want to know about X?" I would ask again and it would answer the same kind of non answer.
> the answer is no it uses the latest gpt models now.
Actually it says it _can_ delegate to the latest models. Seems reasonable to ask how the voice model does when it doesn't delegate (or while waiting for the delegated answer).
I haven't stress-tested it, but I would imagine it approaches complex problems the same way a human with a phone in their pocket would — that is, by having a degree of awareness of the confidence it has in its own knowledge in some areas; where, when it "realizes that it doesn't know", it blocks the conversation with statements like "I don't know, let me check."
I say this because this is already how ChatGPT works internally when using its "auto" mode; the version of the "fast" model used in the "auto" mode does the same "notice your ignorance and bring in the heavy model" thing, just silently, rather than mentioning that it's doing it.
(If someone has actually run the experiment, please chime in!)
> I did report a fun bug with it though: it was interrupting me and laughing at my (not really intended as) jokes while I was still talking! They seem to have clamped that behavior down thankfully, it felt a bit rude and condescending.
Some times I'm still amazed that AI "gets" humour much more effectively than the character of Data did before his emotion chip, e.g.: https://www.youtube.com/watch?v=G4VZ5kQIdV0
Other times I'm amazed in the opposite way, that the script writers cover basically the same talking points about the character as we have today about LLMs, e.g.: https://www.youtube.com/watch?v=GBJCYHwyZhw
I love the UX of voice mode. I’ve always hated voice models.
Gemini’s being particularly egregious (always ending in some deranged question, can not reliably be prompted away) prompted me to build my own client for my real harness that simply does STT -> model -> TTS (both being independently useful).
I guess I see some value in a model responding quickly and with more nuance, but it’s not much. I can wait for it to finish. I’d much rather have it be actually useful. I’m not looking for a digital friend.
The delegation feature lets me see some value in a voice model for orchestration type features. But in either case, I don’t really like (or understand why others would like) talking to a model with different features and quirks just because I’m using a different medium to communicate over.
I’ve spent my time very similarly working on my own voice stack project, but having also seen how non-developers use AI or experience technology in general, I truly think they are better served with a different UX and product than what we have.
In other words, if you’re building your own voice inference tooling you’re just about the polar opposite user demographic than the one that truly needs and will value this. You’re using voice as a medium of convenience doing what existing models are technically and practically “shaped” to be able to do, knowing how they work well enough that conversation is more like typing/prompting with your voice than a natural interface. I’m guilty of this myself but have you ever even paid for a voice/audio model or hardware?
Compare that to the millions of people with an Alexa device in their home who buy products through it, or who prefer calling support to get a human over poring over technical documentation. They’re actually very close to finally getting a version of “Alexa” that lives up to its promise and I’m happy for them
I have built out something similar that let's me use my phone's hardware buttons to open an input stream with the mic for my Hermes agent over my matrix gateway and then has it play back with a local TTS model on my Pixel 10 Pro.
But the Deepseek v4 flash model I am using through OpenRouter is killing me on latency. Any suggestions to improve that?
I have friends who have brainstormed with an LLM (voice chat) for 10-30 minutes, and reported very positive experiences.
When I speak to one - while I'm impressed at how far they've progressed - the LLM just doesn't talk like someone I'd want to discuss a technical problem with (the way I would with a human).
(And my friends aren't even using a custom prompt - some of them are just talking to the default Gemini on their phone!)
I also love exploring ideas like this. The problem that made me stop using Gemini is that it would always try really hard against all my prompting to ask a stupid question at the end of its response that would completely derail my train of thought.
I had an enjoyable conversation with it just now where it repeatedly told me it had generated an SVG of a pelican riding a bicycle along with a clickable data: link. Obviously, this is just harness polish. Overall, it's quite impressive.
Someday we'll have a tokeniser that works directly on human speech instead of doing speech-to-text and text-to-speech steps in between.
>The best feature is that it can delegate questions out to GPT-5.5 in the background, so you're no longer restricted to a voice model that's several years behind the frontier.
Ahh, this makes sense. I was wondering when they would start doing this. I stopped using voice mode all together because it was frustrating talking to a dumb AI, when most of the time I discuss things with Opus 4.8 or gpt 5.5.
I was working on a phone call agent recently, and thought about doing this. It makes sense
I didn't mean the "projects" feature. I described my projects to it, had it search for extra details and pasted relevant inflection into the session before starting the voice chat.
Does it have full access to your chat history, project files, etc? That's the biggest limitation I have with voice mode right now, if I ask it about something I chatted about before (even in the same conversation) it has zero recollection of it.
You can start a chat in an existing session after pasting data into that session and it can then talk about that content. I haven't tried it with projects.
Cool it sounds like they have improved. This the first time I tried back in the day voice chats were different and you could not get the transcript after the fact.
About 2 years ago I used to have conversations with GPT while walking the dog. It really emphasized the need to think before you speak, but you had to think fast before it hung up on you.
Dammit! I live in Europe and the cyber truck isn't even available here..
That said a huge pickup truck is about as far as you can get from a Camaro... Then again I'm not exactly David Hasslehoff myself either... Meh if it talks that's close enough!
That is a nice idea and all, but perhaps there are things our friends aren't good at talking about. Perhaps something that is very related to a work project? Or like when I go down the rabbit hole discussing quantum mechanics and astro physics with ChatGPT because none of my friends understand that at even my level, and I'm looking to learn more about it from someone/something that does know more than me.
I'm looking forward to trying this out because I find that doing discovery on certain things feels more natural with voice. The previous voice chat was such a dumb GPT-4 era model that I resorted to using the dictate feature with GPT-5.5 and had it read the text response back to me.
The best feature is that it can delegate questions out to GPT-5.5 in the background, so you're no longer restricted to a voice model that's several years behind the frontier.
I did report a fun bug with it though: it was interrupting me and laughing at my (not really intended as) jokes while I was still talking! They seem to have clamped that behavior down thankfully, it felt a bit rude and condescending.