AI Engineer Europe 2026
From Transcription to Live Music: Gemini's Audio Stack — Thor Schaeff, Google DeepMind
Read the talk
From Audio Understanding to a Voice-Driven Jukebox
Thor Schaeff moves from a multilingual recording to directed speech, live camera conversation and generated music, showing how Gemini’s audio capabilities fit together.
From a talk by Thor Schaeff
Before you start: Basic familiarity with API requests, JSON schemas and client–server connections is helpful.
Start with a multilingual recording
What can you recover from a recording beyond the words? Thor Schaeff begins by making the test material himself: an introduction that includes his name, a French apology for his French, and a Japanese greeting. He then checks that he has preserved the recording; a previous demonstration had gone wrong when he recorded over it. The introduction will become input to an application that extracts more than a transcript.
Schaeff works on developer experience for the Gemini API and Google AI Studio. His overview follows the DeepMind model families behind these audio experiences. He says he joined in November, the day before Gemini 3 launched. He then introduces the open Gemma 4 family, describing its recent release and opportunities for audio understanding on devices and at the edge. His reference to the previous week belongs to the conference’s April 2026 setting.
Veo, including the cited Veo 3.1 Lite release, brings audio generation into video generation. Gemini 3.1 Flash Live addresses another interaction: full-duplex, sound-to-sound conversation, with real-time text, voice and vision inputs. The rest of the demonstration moves through audio understanding, speech generation and live conversation before combining conversation with music generation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The information a plain transcript leaves behind
Speech carries context, emotion and pacing as well as lexical content. Audio understanding therefore has a broader target than recovering a sequence of words: it must comprehend, transcribe and reason across languages, dialects, accents and modalities. Schaeff also describes Gemini’s ability to handle overlapping speech and switches between languages. Those are capability claims; the upcoming recording provides a narrower multilingual example.
EchoScript puts this into an application. Built with Google AI Studio and available in its gallery, it uses Gemini 3 Flash Preview to analyze a recording and extract information for a structured interface. Schaeff now returns to the introduction he recorded at the start.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
One recording, several kinds of output
EchoScript extracts the analysis in one model request. The summary recognizes that Schaeff introduced himself, and the interface uses that contextual name to label his speech. Alongside the transcript, it presents timestamps, speaker labels, languages and emotions. No audience hecklers join in, so this recording does not exercise the multiple-speaker case Schaeff had hoped to demonstrate.
The results expose both the richness and the fallibility of the extraction. The German passage receives a happy label; Schaeff jokes that earlier attempts had been labeled angry. The prompt also asks for an English translation whenever the speech is not English. His French apology receives a neutral label, rather than the sad label he has seen before. These labels are model interpretations of the performance, not direct measurements of the speaker’s feelings.
The Japanese result misses his intended introduction. The Mandarin result describes a German living in the United States, which Schaeff accepts as correct, though nobody in the audience confirms the Chinese. Keeping these outcomes in view matters: a populated interface can look equally convincing when an individual transcription is right or wrong.
The extraction contract is explicit:
- Identify distinct speakers, using names when the recording supplies enough context.
- Return timestamps and identify the language.
- Translate non-English speech into English.
- Assign an emotion from
happy,sad,angryorneutral. - Put a brief summary of the whole recording at the beginning.
Schaeff reports that one Gemini 3 Flash Preview API call produced the summary and segment-level analysis.
Structured output turns that request into data the interface can consume. Schaeff supplies a response schema, then uses the result to populate the UI. A JavaScript representation of the same output contract can keep the summary separate from repeated segment records:
javascript
const analysisSchema = {
type: "object",
properties: {
summary: { type: "string" },
segments: {
type: "array",
items: {
type: "object",
properties: {
speaker: { type: "string" },
start: { type: "string" },
end: { type: "string" },
text: { type: "string" },
language: { type: "string" },
englishTranslation: { type: "string" },
emotion: {
type: "string",
enum: ["happy", "sad", "angry", "neutral"]
}
},
required: [
"speaker", "start", "end", "text", "language", "emotion"
]
}
}
},
required: ["summary", "segments"]
};
The field layout expresses the application’s needs; it does not make the Japanese result correct. Schema conformance and accuracy remain separate concerns.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Choose a base voice, then direct its performance
The next step reverses the direction: understanding what speech sounds like helps a model generate a particular performance. Schaeff connects the dedicated audio models to Gemini 3 foundational research, emphasizing knowledge of pacing, accents and situations. That connection motivates the speech interface he demonstrates next.
A conventional voice catalog may invite you to filter many recordings or voices by gender, language and accent. Here, the starting point is a smaller inventory followed by performance instructions. Schaeff describes roughly 30 base voices, which can be directed toward different accents and delivery styles. The speech-generation prompting approach treats the chosen voice as a performer whose delivery can change, rather than a fixed accent preset.
The Voice Library application in the AI Studio gallery organizes that direction into several parts:
- Audio profile: who is speaking and what their voice is like.
- Scene: the setting in which the performance takes place.
- Director’s notes: how to deliver the lines.
- Sample context: supporting material for the performance.
- Transcript: the words to speak.
Schaeff requests a high-pitched Irish male voice, then uses Gemini 3 Flash to construct the speech-generation system prompt. It names a character, Finian, places him in a cozy, crowded pub on the coast of County Clare, and asks for a strong Irish accent. The scene gives the delivery a social context as well as an accent instruction.
The live synthesis attempt fails, so Schaeff plays a prepared sample. It tells an animated pub-style story about an enormous thing, a chaotic evening and everyone laughing. He then plays the underlying base voice: a fairly standard American delivery asking what problem they could solve. Replaying the Irish performance makes the change in delivery audible, although the two samples also use different text.
The second comparison uses Sapphire. Its base sample asks whether the listener is ready to build something, again in a standard American English delivery. A Singaporean scene produces dialogue about chicken rice and chili, with expressions including “lah,” “damn shiok” and “confirm plus chop.” Schaeff associates the result with a hawker center. The demonstration combines regional wording with delivery; he attributes that adaptation to the model’s understanding of what the scene sounds like.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A conversation without a text-only middle
Gemini 3.1 Flash Live moves from generating a performance to maintaining an interaction. Through the Live API, a WebSocket connection carries real-time text, audio and video inputs. The model returns audio responses and their text transcripts.
Schaeff cautions against relying too heavily on audio benchmarks and instead emphasizes the architectural distinction:
| Architecture | Where reasoning happens | Response path |
|---|---|---|
| Cascading pipeline | An LLM receives transcribed text | Text response becomes speech |
| Native audio model | Reasoning is integrated into the audio model | Audio response, with a transcript |
A transcript can be an output without being the intermediate representation that drives the conversation. This is the distinction Schaeff draws between the native model and a cascade through text; he supplies no quantitative comparison here.
For an initial trial, he uses AI Studio’s Live interface. At the time of the talk, he describes it as a way to try the models without payment or entering a credit card. The next example tests what happens when audio understanding, visual input and performance instructions operate together.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The camera works—and the accent instruction travels
Schaeff gives the model a system instruction to speak in a friendly Irish accent and enables his camera. He asks whether it can see him, then asks about his outfit. The response identifies his Gemini shirt and backwards hat while maintaining the requested conversational style. Visual input supplies the subject matter; the system instruction shapes its spoken presentation.
He then switches to German and asks for a poem. The model introduces its answer in English and recites German verse. Schaeff hears the Irish accent carry into the German, too. A global instruction has remained active across the language change: changing the language of the request did not narrow the scope of the accent instruction. His recommendation is to adjust the system instructions so the accent does not apply indiscriminately to every language.
Screen sharing is another possible source of visual input. Schaeff states that Live video input is limited to one frame per second at the time of the talk. This is a limit on ingesting frames alongside audio, not a general limit on Gemini’s video understanding. The current Live documentation retains that frame-ingestion limit, but its SDK examples are a later API reference rather than a reconstruction of the on-stage application.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Building the live connection
The implementation starting points follow where the connection lives:
- Python: documentation-linked examples for a server-to-server connection.
- JavaScript: examples for a client-to-server connection.
Schaeff also recommends the published Gemini API coding-agent skills, including guidance for the Live API. His reason is practical: real-time audio adds complexity, and an agent benefits from instructions specific to those APIs rather than having to infer the integration patterns.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Let the conversation request a song
The final component is Lyria 3. A short playback introduces music with sung lyrics, followed by two generation options:
| Model | Output described in the talk |
|---|---|
| Lyria 3 Clip | 30-second jingles |
| Lyria 3 Pro | Full-length songs |
These are generated clips and songs. A live conversational interface around them does not itself imply streaming music synthesis; Lyria RealTime is a separate offering.
Live Jukebox combines the pieces around the familiar experience of calling a radio station to request a song. Schaeff gives the Gemini Live model a tool that invokes Lyria. The conversational model can gather the request, while the music model produces the song. The interface is live even though music generation is delegated to another model.
The DJ persona starts with an on-air greeting and asks what kind of vibe the caller wants. Schaeff interrupts to request German techno schlager about the UK startup scene. The DJ starts proposing a pounding techno beat and a Hans Zimmer-style element; he interrupts again to approve the direction and ask it to proceed.
Before generating, the DJ asks whether the lyrics should include particular startup buzzwords or stories, and whether the energy should be manic or more melodic. Schaeff selects manic energy and leaves the lyrics as a surprise. That exchange supplies a compact music brief: German techno schlager, a UK startup theme, manic energy and open-ended lyrics.
The DJ announces the result, and song playback follows. The German lyrics mention a pitch deck, growth, a Series A goal and coffee, then repeat a refrain celebrating startup life. This completes the demonstrated handoff: a spoken request becomes a clarified brief, a music-generation tool supplies the result, and the application plays it back.
Schaeff closes by offering to rewind to the slides so attendees can retrieve their links—the handoff from the working demonstrations to the resources for building with the same capabilities.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Real-time voice and vision API overview, with WebSocket specifications and Python and JavaScript integration guides.
Installable coding-agent guidance for Gemini APIs, including real-time audio, video, tool calls and session management.
Further reading
Speech-generation examples, available voices, and guidance for controlling scene, accent, pacing and performance.
Examples for generating 30-second clips and longer songs with lyrics, plus prompting guidance and limitations.
- Gemini 3.1 Flash Live model cardDocumentation
March 2026 documentation of Flash Live’s model lineage, modalities, evaluation approach and safety considerations.
Gemma 4 launch announcement with on-device examples and starting points for Google AI Edge Gallery and LiteRT-LM.
Read the complete timestamped transcript
- 0:00
[upbeat music] All right.
- 0:15
Uh, what's new in AI audio? Uh, I'm sorry it's a little bit misleading because the title leaves out the at Google DeepMind. Um, so we're just kind of looking at, you know, what we've been working on at DeepMind.
- 0:28
If we were to look at everything in AI audio, we'd be spending a lot of time here. But, um, yeah, I'd love to show you kind of what we're, what we're working on at DeepMind.
- 0:40
Uh, yeah, this is me. Hi, everyone. I'm Thor. Uh, I work on the developer experience at Google DeepMind, working on the Gemini API and Google AI Studio. Uh, [laughing] Hello.
- 0:52
Welcome. My name is Thor. Bonjour. Je m'appelle Thor. Je suis desole. Mon francais c'est tres mauvais.
- 1:02
Konnichiwa. [laughing] [speaking foreign language] Okay. Uh, that was for the demo, and now I just need to make sure... Last time I did this demo, I recorded over it, and then it was all gone.
- 1:21
That was very sad. But we'll, we'll come back to that in a bit.
- 1:26
Yeah. What have we been up to at DeepMind? Um, there's been a couple releases. Uh, I, I actually joined the team in November, uh, literally the day before Gemini-Three was released.
- 1:38
So I joined, and they told me, "Tomorrow we're releasing Gemini-Three." And I was like, "Yay." Didn't do anything, but it was great. Um, it was, it was a great time.
- 1:48
Uh, most recently on the open model side, we released Gemma-Four, uh, I think literally last week. And, um, yeah, pretty incredible. Some cool stuff you can do there. Multimodality as well, uh, baked into Gemma-Four, so, uh, there's audio understanding in, uh, the, the Gemma-Four models, and you can do that, uh, on device, on kinda edge devices as
- 2:10
well. So, uh, that is some, some very exciting stuff. In terms of, uh, Genmedia and audio, uh, you're probably very familiar with our, um, you know, image generation models, video generation models.
- 2:23
Obviously Veo has audio generation in there as well. Um, so this is kind of where the progression, uh, is there most recently with, uh Veo Three Point One Light on the Genmedia model site.
- 2:36
And then on the audio models, um, we recently launched Gemini Three Point One Flash Life, which is our kind of full duplex, um, you know, sound to sound, uh, real-time conversational model.
- 2:50
Also multimodal, so you can ingest, uh, real-time text, voice, vision, um, which we'll, we'll look at in a, in a bit.
- 3:01
So, you know, on kind of audio, very, very broad topic, but sort of the, the baseline of everything we do are kind of the, you know, frontier Gemini models.
- 3:10
And so Gemini Three is, um, incredibly good at understanding audio and, you know, that's not just transcribing it, uh, but really, you know, understanding kind of all the nuances that are in there.
- 3:23
So that might be, you know, obviously speech, but also the context of the speech, the, the emotion, um, you know, your pacing, your, um, sort of... Yeah, a-anything that sort of swings within the audio that's, you know, not just text.
- 3:39
Um, so on the audio understanding, kinda our goal is to build models that deeply comprehend, uh, richly transcribe, and robustly reason through audio, uh, seamlessly handling, you know, a large mix of different languages, dialect, accents, and modalities, and sort of anywhere and always.
- 3:58
You know, Gemini is really good as on, you know, transcribing even people that are talking over each other, um, which is, which is pretty, pretty incredible. You know, seamlessly switching between different languages.
- 4:10
That was sort of the demo we're looking at, um, now. So, um, EchoScript is kind of Gemini 3 Flash Preview to sort of analyze, you know, audio recordings and extract sort of information out of it.
- 4:25
Uh, it is built with Google AI Studio, so you can, you can find it in the gallery in AI Studio. You can try it out. Uh, I can give you the slides later as well.
- 4:35
Um, and so that was sort of, you know, what I was trying to demo earlier.
- 4:40
Um, so, you know, different from just kind of a pure transcription, uh, model. W-we can extract a lot of information out of the audio within one single, you know, request to the model or one single API request if we're using the API.
- 4:57
So you can see here, you know, summary. I introduced myself by name. Um, so we're actually able to label, um, you know, the section with speaker by name. Uh, I forgot there was no hecklers in the room, otherwise we would have picked that up as well.
- 5:10
Um, and maybe we can see if we have more time later and we can do that. But so we can see here, you know, we're extracting timestamps, uh, we're labeling the speaker, kinda identifying the speaker, we're, we're identifying the language and sort of the emotion of it, right?
- 5:24
Uh, happy, you know, to introduce myself, uh, is great. Um, now you can see this was in [REDACTED:origin]. Um, normally, it would classify my [REDACTED:origin] as angry. [laughing] Uh, but, uh, here, you know, I, I guess I'm very happy to be with y'all.
- 5:39
So, uh, I just told it, you know, uh, label sort of the emotion, uh, label the language. If it's a language that is not English, um, give me an English translation as well, right?
- 5:51
Um, in French, uh, neutral. Normally would say sad. Uh- [laughing] You know, French is just a bit more of a... Uh, no, I sat... You know, I'm sorry, my French is very bad.
- 6:02
Uh, didn't, didn't sound sad enough, so neutral in this case. Um, [laughs]
- 6:09
okay, this, uh, this didn't work, so my Japanese, I gotta, gotta practice that if anyone reads Japanese. So it should actually say, "Hello, my name is Thor." Unfortunately, bit of a miss there.
- 6:20
Uh, let's see if my Mandarin was any better. "Hello, everyone. I'm a [REDACTED:origin] living in the United States." Sorry, my Chinese... Yeah, that is, uh, correct. Does anyone read Chinese in the room?
- 6:30
No. Okay. Well, uh, we'll just trust that that is, um, that is correct. Uh, and so we can see here that, um, this was, you know, one, uh, kind of request to the model, um, where, you know, I basically just told it, "Identify the distinct speakers.
- 6:47
If you have context, you know, label it by na- label the speakers by name. Give me the accurate timestamps. Give me the language. If the language is not English, give me the translation.
- 6:58
Uh, identify the emotion out of, you know, happy, sad, angry, neutral, and then also provide a brief summary of the entire audio at the beginning." So this was one API call, uh, to Gemini 3 Flash Preview, and you know, we got all this information out.
- 7:12
We could, you know... I just gave it a response schema, so kind of structured outputs, uh, and I was able to just populate that into my, into my API, uh, into my UI to have sort of the structure.
- 7:24
So you know, this kind of audio understanding and sort of the, the, you know, base research in the Gemini 3 models, that is sort of what powers, um, the speech generation as well, as well as the, um, you know, real-time conversational, um, generation.
- 7:41
So having that audio understanding, um, is, is really, really great in terms of, you know, knowing what certain things sound like, um, including, you know, different pacing, different accents, and, and, and scenarios like that.
- 7:55
So, um, you know, the foundation of kind of all our models is now sort of the, the Gemini 3, um, foundational research, and then we're building kind of the, um, dedicated audio models on, on top of that.
- 8:07
And so with speech generation, uh, it's a bit different. You know, if you've used kind of other, um, TTS providers before, uh, you probably have a huge library of, you know, different voices that you sort of, you know, you filter by gender, by, by, you know, accent, by languages, what have you.
- 8:25
Um, but so in, in, in Gemini, you have, you know, just I think it's, like, 30-odd sort of base voices. A- and then what you do is you, you kind of direct that voice to act in a certain way.
- 8:38
Uh, and again, because we have that audio understanding, uh, we can, we can basically modify the voice to, you know, act in a certain way to, you know, use a certain accent.
- 8:49
Uh, and so we can go from kind of a small set of, of base voices to a very specific, you know, kind of voice that we're looking for for our speech generation.
- 9:01
Uh, again, there's, um, a little application that you can try out. Uh, it's in the, uh, Google AI Studio Gallery as well. Uh, it's called the Voice Library.
- 9:14
Uh, and so what we can do is, you know, um, kind of giving the, the, the prompt structure that we just saw, you know, we're building sort of the, the audio profile, the scene.
- 9:23
We're setting the scene. We're instructing sort of this director's notes, so we're giving guidance for the performan- you know, just like how you would, um, direct a human, you know, to, to act out a certain way.
- 9:35
Um, and then some sample context and kind of the transcript, um, that we want. So now what we can do is, you know, we, we just said sort of, uh, we want, you know, a high-pitch Irish male.
- 9:47
Uh, and so basically, I just used Gemini, um, 3 Flash here again to then construct our, um, system prompt for the speech generation. So we're saying here, you know, um, we're, we're setting sort of our au- audio profile.
- 10:02
You know, Finian here, and the scene, sort of cozy crowded pub, uh, in the coast of County Clare. Um, you know, deliver the lines with a strong, authentic, uh, Irish accent.
- 10:14
And so now we hope the, the TPUs don't, uh, disappoint me. There we go. It failed. But, um, I didn't, you know... I prepared it, so we can, we can listen to it here.
- 10:25
Oh, you wouldn't believe the size of the thing until you saw it with your own two eyes, I'm telling you. It was a grand old mess, so it was, and we were all laughing fit to burst by the end of the night.
- 10:36
So as, as you can see, you know, this was, uh, the, the base voice here is, uh, this one.
- 10:42
What kind of problem could we solve?
- 10:45
So you know, that is a fairly sort of stan- standard, you know, American accent, but so now by, you know, giving it that director's note, we can then sort of give-
- 10:55
Oh, you wouldn't believe the size of the thing until you saw it with your own two eyes, I'm telling you.
- 10:59
Takes us straight back to Dublin.
- 11:00
It was a grand old mess, so it was, and we were all laughing fi-
- 11:04
Or, you know, similarly here we have, um, Sapphire. So this voice, uh, is here.
- 11:10
Ready to build something awesome today?
- 11:12
Again, you know, kind of fairly standard sort of American, um, uh, English accent here. And now we could say, you know, give it kind of a Singaporean sort of scene.
- 11:22
Wow, you must try this chicken rice, love. The chili is damn shiok. Confirm plus chop you will love it. Faster queue before the uncle close shop, okay?
- 11:31
Anyone spent time in Singapore? Uh, you know, yeah. That's, uh, you know... That's something you would hear in the Hawker Center. So, um, again, you know, that is kind of underpinned by the audio understanding.
- 11:43
So, um, the model really understands what, you know, these, these different scenarios sound like and then can modify, uh, the speech generation to, to, to be like that.
- 11:55
Um, yes. And then, you know, finally- Sort of the, the native audio, uh, sound to sound multimodal real time. So we, uh, just launched a couple weeks ago Gemini 3.1 Flash Life.
- 12:09
Um, so it is, uh, a speech to speech, this kind of real-time multimodal model. Uh, you can ingest, uh, text, audio, video in real time through, uh, a ve- WebSocket connection, and then you get, uh, real time audio response back as well as kind of the text transcript of that.
- 12:28
Um, you know, obviously benchmarks are, you know, especially in the audio space, benchmarks, you know, you can't really trust them. Um, you know, it's great. Y- you can see sort of the reasoning, the thinking.
- 12:40
Uh, so, so here the, the, the thinking and the reasoning and the intelligence is baked directly into the model. So that's, you know, different from a cascading pipeline where you would actually go, uh, through text to then go through an LLM to get the intelligence.
- 12:56
Here, you know, the intelligence is baked into the audio model, so that's kind of the difference there. Um, but obviously in, in real scenarios, um, you know, you can try this out in, uh, ai.studio/life.
- 13:09
So the great thing with AI Studio is, uh, you can try it out, um, you know, without paying anything. So this is, uh, allows you to, you know, try kind of, um, the models without actually needing to put down the credit card.
- 13:24
So AI Studio is a, is a really great way to do that. Uh, and again, you know, we have the audio understanding kind of baked into the model here.
- 13:32
So what we can do is we can give it some system instructions, you know, for example, speak in a friendly Irish accent, and then also we can ingest kind of our, you know, camera here, for example.
- 13:43
And then, um, we can say, "Hey, can you see me?"
- 13:48
Well, hello there. I can see you loud and clear. So I can. What's on your mind today?
- 13:53
What do you think of my outfit?
- 13:56
Ah, look at you with your Gemini shirt. It looks grand altogether, and the backwards hat gives you a fierce, laid-back vibe, so it does. You're looking sharp. [laughs]
- 14:07
Uh, [laughs] Wunderbar. Kannst du mir ein Gedicht auf Deutsch erzählen, bitte?
- 14:12
Ah, a poem in [REDACTED:origin], is it? Sure, I can give that a go for you. Here's a little one. Es war einmal ein kleines Licht, das strahlte so hell und mit, es brachte Freud in jedes Herz, vertrieb den Kummer und den Schmerz.
- 14:28
How was that?
- 14:29
So what's funny here now is, um, it actually applies the Irish accent to the [REDACTED:origin] as well. So obviously you need to kind of adjust your system instructions to, you know, not speak in an Irish accent in any, every language.
- 14:43
Um, but so it, it, it's pretty funny. You can kind of switch between, um, the different languages there. Uh, again, so ai.studio/life. You can try it out. Uh, you could also ingest your screen.
- 14:55
So you're basically just ingesting, um, video frames in addition to the audio at a maximum frame rate of one frame per second at the moment. That's, that's kind of the maximum you can ingest.
- 15:09
Um, there is, uh, a couple examples, you know, if you're a developer yourself, uh, feel free to try these out. They are linked from the docs as well. These are, you know, just some examples of how you would get started, say, in Python for like a server to server connection or in JavaScript with kind of a client
- 15:25
to server connection. So you can find those here. Uh, also recommend the, uh, Gemini skills. So we have published coding agent skills for kinda all of the Gemini APIs, including the Live API.
- 15:37
So, you know, real time audio, working with real time audio can just be a bit more challenging. So, uh, using these agent skills and kind of installing them in your, uh, you know, coding agents can really help steer them sort of in the right way and give you sort of that result that you're looking for.
- 15:55
Um, lastly, we have... Okay, we have a bit of time. So music, uh, you know, also audio, so I put that in here. Uh, we recently released Lyria 3.
- 16:09
Silver glow and a steady beat. Again and again.
- 16:18
So yeah, it's, it's a music generation model. Um, but so it, it now actually can generate music with lyrics. There's two separate models. There's a, uh, Lyria 3 Clip, uh, which is a 30-second kind of jingle gener- generation model, and then a Lyria 3 Pro is the full length, uh, song generation model.
- 16:39
Um, and so I, you know, kind of bringing everything together, I don't know if you remember the good old days when you would actually call the radio station and wish for a song.
- 16:49
Um, and so I've, I've kind of built this application, um, called Live Jukebox, and so we can now try it out where, um, we actually give the, the real time Gemini Live model, uh, a tool to then generate a song using Lyria.
- 17:06
And so let's see if we can, if we can do that here.
- 17:09
Right then. You're live on air with the Jukebox DJ. What an absolute legend. Who am I speaking to? And what kind of vibe are we gonna cook up today?
- 17:17
Is it a proper rave starter or something a bit more ch-
- 17:19
Yeah, can we get, um, maybe a [REDACTED:origin] techno schlager about the UK startup scene?
- 17:28
[REDACTED:origin] techno schlager about the UK startup scene? Mate, that is revolutionary. I love it. Right. Picture this. Pounding techno beat, a proper Hans Zimmer style b-
- 17:38
Yeah, that's great. Let's go. Let's do that.
- 17:41
Hold on tight. We're almost there. Just to make sure it's a certified banger, any specific buzzwords or stories from the UK startup world you want in the lyrics? And are we thinking manic energy or something a bit more melodic within that techno madness?
- 17:54
No, just manic energy and surprise me on the lyrics.
- 18:01
All right, you're on. Get ready to have your eardrums blasted. Manic [REDACTED:origin] techno schlager with a British startup twist. Cooking up a proper banger for you. Check this out.
- 18:11
Zeig mir dein Pitch Deck los. Der Wachstum ist grandios. Series A ist jetzt das Ziel. Das Spiel ist ein Gewinner-Spiel. Viel Kaffee in meinem Blut. Das Startup Leben ist so gut.
- 18:30
Das Startup Leben ist so gut. Ja, ja, ja, so gut. [laughs]
- 18:41
All right. I'll, I'll leave you with that. Um, if you don't speak [REDACTED:origin], sorry, but, um, thanks so much. Appreciate y'all. [applause]
- 18:50
No worries at-
- 18:55
Oh, yeah, and if you want the slides, uh, I can just rewind, but there's, like, all the links in there. Um, if that hel- if that's helpful, you can, you can just grab them there.
- 19:05
Awesome. Thank you. And yeah, enjoy the rest of the conference. And big thank you as well to our friends in the back on the audio. You know, wouldn't be possible without them.
- 19:14
Cheers. [applause] [upbeat music]