Gemini Live Avatars — Transcript
Full transcript
- 0:00Welcome. I'm Jennifer. I'm ready to dive
- 0:02into the latest F1 action with you.
- 0:05>> Hey Jennifer, who won the latest Grand
- 0:06Prix?
- 0:07>> George Russell won the latest Grand Prix
- 0:09in Azerbaijan. He had a fantastic race,
- 0:12holding off Max Verstappen to take the
- 0:14win.
- 0:14>> So, where did Kimi Antonelli come?
- 0:17>> Kimi Antonelli finished in fifth place
- 0:19after a really impressive recovery drive
- 0:21from 16th on the grid.
- 0:23>> Okay, so that was Google's new live
- 0:26avatars with Gemini 3.8 live. So, I've
- 0:29talked about the live models in the
- 0:30past. These are the bidirectional models
- 0:32that basically have multimedia going in,
- 0:35being audio and video. And up until
- 0:37recently, it could just produce text or
- 0:39audio back. Well, Google's now updated
- 0:41that so it can actually have a live
- 0:43rendering face that kind of lip-syncs
- 0:46what the actual audio is. So, up until
- 0:48recently, if you wanted to do any sort
- 0:50of AI agent with a face that could
- 0:52actually hold a conversation, you would
- 0:54generally sort of stitching together
- 0:56four different things. You had some kind
- 0:57of speech to text, you had an LLM, you
- 1:00had a TTS model creating the voice back,
- 1:03and then there was some sort of avatar
- 1:04service on the end doing lip-sync. And
- 1:06the challenge has always been with that
- 1:08kind of thing is that every one of those
- 1:10points adds a lot of latency. And by the
- 1:12time the mouth starts moving, the
- 1:14conversation has often already started
- 1:16to fall apart. Now, I won't say that
- 1:18Google's got this perfect, but what
- 1:19they've done is pretty impressive.
- 1:22They've made their Gemini 3.8 live
- 1:24avatar system generally available. And
- 1:26there are actually a few caveats to
- 1:27that, which I'll mention later on in the
- 1:29video. But this whole thing basically
- 1:31collapses down those four models that
- 1:33you had to take care of yourself into a
- 1:35single stream. It can listen, it can see
- 1:38what's on your screen, it can call tools
- 1:40in the background. As you can see in the
- 1:42example I just showed, it was using
- 1:44Gemini search to basically get the
- 1:46latest results from Saturday's Grand
- 1:48Prix. And the cool thing here is it can
- 1:50do all of this in 97 languages. So, I've
- 1:53been playing with this for a while, and
- 1:55I built some demos that I'm going to go
- 1:56through in a second and show you. But
- 1:58let's put this into a little bit of
- 1:59context. So, while we're still yet to
- 2:01see a Gemini Pro model come out now for
- 2:04I think it's about 7 months, this has
- 2:06still been a pretty busy week for Google
- 2:08on the audio side. So, in my last video
- 2:10I talked about the whole new TTS system
- 2:12that they launched with the voice
- 2:14cloning, with those sorts of things. And
- 2:16actually before that, they updated
- 2:17Gemini 3.8 live and live extended
- 2:20thinking. And now they've released live
- 2:22avatar. And like I said in the last
- 2:24video, whatever you think about Google
- 2:26is in the LLM race at the moment, the
- 2:28multimodal and audio stuff has been
- 2:31really strong. They've had Omni, live
- 2:33translate, transcribe, the TTS models,
- 2:36and now this live avatar. So, I guess
- 2:38quickly what does generally available
- 2:40actually mean here? Basically, it's
- 2:42ready for production on Google Cloud.
- 2:45But the sad thing here is that not
- 2:46everything is generally available. You
- 2:48get a library of pre-built avatars. If
- 2:50you want a custom avatar, one made from
- 2:53your own reference photo, your own sort
- 2:55of voice, that kind of thing, that's
- 2:56behind an allow list and a verification
- 2:58process. And even things like the live
- 3:01extended thinking version of live is
- 3:02still in private preview. So, in this
- 3:04video I'm just going to show what you
- 3:05could go and do and build right now and
- 3:08show that in my demos, etc. If you want
- 3:10to see an example of the custom agents,
- 3:12they've got a video in there with a demo
- 3:14of making a custom avatar for an avatar
- 3:17with a really bad Australian accent that
- 3:19I think only a bank owned by Crocodile
- 3:21Dundee would actually want to use. So,
- 3:23with what they've actually made
- 3:24generally available in here, you've got
- 3:27a bunch of different visual avatars that
- 3:29you can pick from. So, here I've made an
- 3:30avatar studio, which is basically a
- 3:33front end app that's running this and
- 3:36streaming the live avatars through the
- 3:37back end in here. So, here we can
- 3:39basically just pick any of the avatars
- 3:42that we want to. I can then basically
- 3:44choose a voice that I've got and they've
- 3:46got a set of voices that we can actually
- 3:48pick. We can pick the sort of persona
- 3:51preset that we want. Let's go for a time
- 3:53traveling tour from 2126, and then we've
- 3:56got the system instructions. Now,
- 3:57obviously, the system instructions, you
- 3:59can fully customize yourself. You can
- 4:01get all of that set up. And then you can
- 4:03add in tools if you want. So, here I've
- 4:05basically got Google Ground Searching
- 4:07for this. Now, since we're talking to
- 4:09someone from the future, that actually
- 4:10might not be good. I will leave that off
- 4:12for now. And then basically, what I can
- 4:14do is I can just start it. And then I've
- 4:17got the choices. Once I start
- 4:19>> Wow, it is so incredible to finally see
- 4:22the quaint 21st century in person.
- 4:25>> Okay, and you can see there that now I'm
- 4:28talking to Vera. Vera, how are you doing
- 4:30today?
- 4:31>> I am doing absolutely wonderful. Thank
- 4:33you. Exploring this era is just pure
- 4:36magic. These are actually a perfect
- 4:39vintage replica of 2026 fashion. So, you
- 4:42could say I dressed the part to blend in
- 4:44perfectly.
- 4:45>> Okay, so Vera, what I want to show is
- 4:46that we can actually type to you as
- 4:48well. So, I'm going to ask you a
- 4:50question in a second, and I only want
- 4:52you to answer the question when you see
- 4:53the typed question.
- 4:55>> Understood. I am ready and waiting for
- 4:57your typed question. I'm doing
- 4:59absolutely fantastic. Just so clear the
- 5:02incredible energy of 2026.
- 5:06>> You can see there that basically, she
- 5:07was able to take a type or voice, and
- 5:10she's still listening to me, aren't you,
- 5:11Vera?
- 5:12>> Exactly. I'm processing everything
- 5:14perfectly, no matter how it reaches me.
- 5:16>> Okay, I'm going to stop there. So, you
- 5:19can get a sense from that this is
- 5:20actually pretty responsive, right? Now,
- 5:23there is a delay, and obviously, we're
- 5:25going to get the transcript back ever so
- 5:27slightly before the actual voice comes
- 5:30back, which kind of makes sense. But you
- 5:32can see for something where now we can
- 5:33go through and pick any of the different
- 5:35avatars. You can pick photorealistic
- 5:38avatars, or you can pick sort of
- 5:40cartoonish avatars in here. And so, the
- 5:43cool thing here is you can actually make
- 5:44a bunch of fun demos with this. So, not
- 5:47only can we do all of this in English,
- 5:49we can actually come in here and we can
- 5:51pick different languages. So, if we want
- 5:54to have a cafe and we want to put it in
- 5:56a different place, we can actually sort
- 5:58of make this a conversational app for
- 6:01teaching us to speak a different
- 6:03conversation, etc. So, you can see here
- 6:05we've basically got something where it's
- 6:08going to teach me Let's try Japanese.
- 6:11Let's see how that goes. And what I'm
- 6:13going to do is try and have a
- 6:14conversation with her. Hi there. Can you
- 6:16tell me how to order a coffee in
- 6:18Japanese?
- 6:25Kohi is coffee?
- 6:31Okay, so how do I say I would like one
- 6:33coffee?
- 6:38How do I pronounce it? Can you give me
- 6:40the corrections for how I should
- 6:42pronounce it?
- 6:50Kohi o kudasi.
- 6:56Okay, so enough of me embarrassing
- 6:58myself Nothing going to speak Japanese,
- 7:00but you can see here that we can
- 7:02basically set this up to teach us
- 7:05vocabulary, to do a whole bunch of
- 7:06different things. And we could change it
- 7:08so that we start with something that's
- 7:10really beginner and something that then
- 7:12gets right through to something much
- 7:14more advanced in here. Another thing
- 7:16that you can actually do for fun is that
- 7:18you can actually have multiple
- 7:20characters and have them communicate to
- 7:22each other. So, here we're going to
- 7:24start a debate over whether pineapple
- 7:27belongs on pizza or not. And you can see
- 7:30Moderator, esteemed guests, culinary
- 7:33purists often mistake rigidity for
- 7:35tradition, but love and gastronomy
- 7:39thrives on daring unions. To deny the
- 7:42pineapple its rightful throne upon the
- 7:44pizza is to deny the very spirit of
- 7:46romance, which seeks the unexpected
- 7:49harmony of sweet and savory. Imagine the
- 7:52sun-drenched tropics marrying the rustic
- 7:54warmth of baked dough. It is not just a
- 7:57topping, Carmen. It is poetry on a
- 7:59crust.
- 8:00>> Quickly, it is then. Leo, while I
- 8:03appreciate your poetic vision, culinary
- 8:06harmony requires more than just sweet
- 8:09and savory colliding. From a practical
- 8:11standpoint, the excessive moisture in
- 8:14pineapple compromises the structural
- 8:16integrity of the crust, turning a crisp
- 8:18base into a soggy mess.
- 8:21>> You are right, moderator. There is
- 8:23wisdom in her words that I must respect.
- 8:26Yet, Carmen, true passion often
- 8:28withstands a few
- 8:30>> Okay, I'm going to stop it there, but
- 8:31you get a sense there of that these two
- 8:33can actually sort of communicate to each
- 8:35other. So, you can have some sort of fun
- 8:37with that. Probably I'd go back and
- 8:39actually change the system prompt to
- 8:40make them a little bit more concise for
- 8:42my demo. And it doesn't stop there. You
- 8:44can also do things like where because
- 8:46you can also activate the camera, you
- 8:49can then hold up things and have a
- 8:50conversation about them with the avatar.
- 8:53Now, if you want to see a video of me
- 8:54actually explaining the code and going
- 8:56through it a little bit and stuff like
- 8:58that, I showed you how to get started,
- 8:59etc. I'm going to put the link for that
- 9:01in the description. That's going to be
- 9:02on my second channel. You can go through
- 9:04and get this all started yourself. Okay,
- 9:07so just to finish off, I would say that
- 9:08while this doesn't have the flexibility
- 9:10that perhaps we would like at the
- 9:12moment, again, this is Google having to
- 9:15be conservative with how they do things.
- 9:17They don't want 500 people uploading
- 9:20pictures of Trump or making a Trump
- 9:21avatar or something like that. My guess
- 9:23is there's going to be quite a number of
- 9:24safety features around that kind of
- 9:26avatar system. You can imagine also that
- 9:28Google's going to be pretty strict about
- 9:29the different use cases for this. If you
- 9:31want to try it out, this is in Google
- 9:33Cloud now. You can actually just come in
- 9:35and look at the Gemini live avatar. You
- 9:38can pick from the avatars that they've
- 9:39actually got set up in here. So, you can
- 9:41certainly get a demo of this up for a
- 9:43business or to show people how to use
- 9:45it, etc. Now, the pricing for this is
- 9:47going to be pretty pricey coming out. I
- 9:49think the simple fact that this just
- 9:51uses so much compute means that we're
- 9:54probably not going to see this get super
- 9:55cheap in the really near term. But, it's
- 9:57definitely going to be interesting to
- 9:58see where this goes, and I'd love to
- 10:00hear from people where you see the best
- 10:02use cases of this are actually going to
- 10:03be. Let me know in the comments what you
- 10:05think about those. And hopefully it
- 10:07won't be too long before we see an open
- 10:08version of this, perhaps from Nvidia or
- 10:11from one of the other companies that are
- 10:12actually releasing open models out there
- 10:14as well. Anyway, as always, if you found
- 10:16the video useful, please click like and
- 10:18subscribe, and I will talk to you in the
- 10:19next video. Bye for now.
About this transcript
This page contains the full transcript of Gemini Live Avatars by Sam Witteveen, generated from the public captions YouTube serves with the video. The transcript has 2,012 words across 289 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.