Speech-to-Speech Model Research at Google DeepMind — Valeria Wu Fon & Tom Ouyang, Google DeepMind

AI Engineer · 16 min · 193 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:01[music]
00:12Hi everyone.
00:13Thanks for coming to our talk.
00:14We're going to talk a lot about how we are planning on voiceifying the agentic future with speech-to-speech research here at Google DeepMind.
00:22My name is Valeria.
00:23I'm the product lead for the speech-to-speech model in Gemini and Tom.
00:27Yeah, my name is Tom.
00:28I work on speech-to-speech as an engineer and Gemini.
00:31So, yeah, great to be presenting our work.
00:33Cool.
00:35So, before we start diving into the history of Google research and audio and the latest research we've been doing, I just want to take a step back and really kind of like think about why this is important to us in the team.
00:47And the main reason is that voice is the most natural way for humans to interact with both the physical and the virtual world.
00:53And we're already seeing so many applications that are starting even within the own Google products.
00:58We see voice being used to ask questions and you know, do homework or info seeking or EDU in search live and Gemini live.
01:05All the way to these same models being deployed in cloud and the API for enterprise or voice agent use cases.
01:11Should I use this as a microphone?
01:15Hello.
01:16Okay.
01:16Um So, because of that, we only think that the number of applications is going to, you know, exponentially increase over the next few years and we believe that speech-to-speech models are the way to go when we want to build robust universal voice agents.
01:30And with that, I'll hand it off to Tom.
01:33All right.
01:33So, of course, like one way to create a voice agent is to just add an ASR like speech-to-text model in the pipeline and then a text-to-speech model on the other end, right?
01:41So, of course, that's where a lot of the history of speech come from.
01:44Of course, Google has been working on this for a long time.
01:46I'm going to give you a little bit of a historical overview of what speech modeling, especially speech-to-text automatic speech recognition, typically looks like, right?
01:54So, up until around 2018, this usually involved a lot of different components.
01:58You have um you know of course feature extraction that's fairly general and then you have all these different pieces like acoustic modeling, pronunciation modeling, language modeling, a second pass rescoring
02:08that allows you to go from the audio input to a text transcription.
02:12Right?
02:12And of course like around 2018, these moved more and more towards end-to-end systems.
02:15You don't have to have so much domain knowledge.
02:17You can actually have mostly the the neural model learn this pattern and mapping between acoustic inputs and text.
02:24But these aren't really the end-to-end models that we think about when we think about LLMs.
02:27They're only doing kind of one thing, which is speech to a transcription of that speech.
02:31They're not responding, they're not translating.
02:33If you wanted to have the model tell you about the tone or the emotion or the speed of the speaker.
02:38If you wanted to bias it towards words or much less like images, those are all things you have to build yourself as part of the system and there's really really a barrier to how easily you can scale
02:47these systems.
02:48So, fast forward to now, which is kind of of course the LLM era, right?
02:52Like you know, the first LLMs were mostly text, but even there I think you could kind of hack audio embeddings into these text models and it kind of worked, but now of course for a long time now, Gemini models have been very natively
03:02multimodal.
03:03So, what does that mean?
03:04It means that when we train these models in pre-training, which is where the bulk of the data comes from, these are multimodal interleaved examples, right?
03:10So, the bottom uh diagram here gives you kind of one example of what that might look like.
03:15So, this is a task where you're asking the model to summarize a bedtime story and there's a text prompt in the beginning, but then there's this sequence of video and audio
03:24inputs that the model gets and then of course like what you expect the model to do here is produce um both the summary and also annotate timestamps for where interesting things happen and so forth.
03:33So, this example is teaching the model to translate its understanding of the audio and the video into text.
03:39You might have other examples in pre-training that ask you to caption a video.
03:43So, you might have a video that has audio and the model is learning to bias towards both the video and the audio signal to caption this well.
03:51And of course like there's limitless like YouTube videos with captions that you can train these models on.
03:55Other models might actually try to generate audio from the video from the text, right?
03:58So, you can have ASR, TTS, or any combination of these plus all of these sort of agentic tasks all kind of learned under one unified token embedding space.
04:08So, this becomes a foundation for a lot of what we want to do in audio because we already have a model that understands audio, video, text, and how these things relate and transition from one to the next.
04:19So, very quickly, like one of the applications that this enables that we've launched recently is live translation, right?
04:25And this is a kind of application that kind of only works when you have all of these capabilities working within the same model.
04:30You have basically state-of-the-art translation quality.
04:32Even though this model is translating basically as the user or speakers are speaking, you kind of ask, "Hey, I I speak English.
04:38There's maybe friends who are talking in Spanish and Italian and Chinese."
04:42And it's translating all of them to your language as they talk.
04:44Um and then we're we're finding is the translation quality for this like streaming real-time translation is about as good as you would get with offline systems, right?
04:52Where you kind of know the full utterance um from the very beginning.
04:55So, that's something that has been classically very hard to do with these cascaded systems, but with LLMs, it actually just a lot of it comes out of the pre-training.
05:05So, of course, to do this task it needs to do multilingual switching because you could be translating across different languages.
05:09You don't know what those languages are beforehand.
05:11It needs to preserve the speaker voice of the source speaker and be able to understand multiple speakers, be robust to noise, and of course, like do all this in real time, right?
05:21So, again, it would be very hard to try to engineer this, but then with the LLM and Gemini models, this almost becomes a prompting task.
05:29And on that, like, I know, in this diagram we're saying, "Hey, at the top with these models, if you prompt it to do the speaking like this streaming translation task,
05:38and you give it the audio, it will produce the streaming translation output, right?
05:41The same model, if you ask it to act like an agent and respond to maybe image and audio input, maybe asking questions about that image, it will give you an answer.
05:50And finally, like very well we'll show examples of this, you can also have it create this embodied, you know, virtual agent that has a face, that has, you know, things and tools that it can show you, and it will produce this sort of embodied agent experience.
06:02So, with that, I'm going to give it to Valeria to talk more about the North Star and some of the key demo products that we built.
06:10Yeah, so to create this type of kind of universal, versatile, uh kind of model/product, there's like three vectors that we think about when we do research and product for these models.
06:20Um and they also come with some challenges, so I'll like walk you through some of them.
06:24So, at the core of a speech-to-speech model, the first thing that people usually think about is that it has very conversational, right?
06:30It's low latency, it's very conversational, very snappy, very natural.
06:34But, I think within our team, we really don't only want this model to sound nice.
06:38We also have two pillars at the top that we also really care about like pulling all together into one model, which is intelligence and it being multimodal.
06:46So, when we talk about intelligence, we talk about, you know, task completion, instruction following, reasoning, like capabilities that the model needs to have natively in order to complete tasks and to like do things
06:57uh that have high customer satisfaction, for example.
07:00And on the other side of the Venn diagram, we also have the idea that these models should be very multimodal, both in audio in and audio out, right?
07:07So, sometimes a user doesn't only want to input audio in and have that be the start of the conversation.
07:12We need video, your screen sharing, PDFs, whatever you would want the model to interpret and understand, we should be able to stream it in and also produce output out of it.
07:21So, that's kind of like the trifecta of which we think about speech-to-speech models.
07:26Um and I want to add a caveat about ITNN.
07:28I think actually the majority of our Gemini users are non-English speakers.
07:32Um so, we put a big focus on having and making sure that all these capabilities work within not only, you know, EN-US, but all the languages that our customers care about.
07:41Um of course, this also our North Star also becomes one of the biggest challenges in our research because once you move one of the knobs, it's very easy for the other knobs to kind of like mess up, right?
07:52Like a very quick example, oh, how do we increase intelligence in the model?
07:57Well, you can turn thinking high or like the thinking is high as possible to have the model think a lot before calling a tool or answering a question, which in eval's it does show that it does improve the model's intelligence.
08:08But when what does that do to latency, right?
08:10And time to first audio and the naturalness of the conversation?
08:13So, within the Gemini team, we're really trying to push forward research initiatives that can kind of blend in the three of them without really sacrificing any of those by a lot.
08:23Um, but in the meantime, we're going to show you some of the demos that we think are hinting at how our speech-to-speech model can combine all these three into
08:32really cool application.
08:33So, the first one is you kind of saw this as a preview, but our live model, as you know, uh, powers streaming translation that supports 70+ languages.
08:41So, this is a little bit around the core of conversation quality and IT&N efforts that we have in the team.
08:47I'll play a quick video on how this works on Google Meets to help two people or maybe multiple people that are speaking different languages still have a live conversation.
08:57Oh.
08:58One sec.
08:59Okay.
09:00Okay, let's turn on speech translation.
09:04So, here the user can simply select the language that they want the translation to happen in and then the rest will be done in real time in multiple speakers kind of having a conversation back and forth.
09:14So, I'll show you a snippet of this video of what happens.
09:20It's great to see you both.
09:22Cassie, how's the weather in Shanghai?
09:25It's nice to meet you.
09:26The weather here is really nice, sunny and bright.
09:29I spent all day Saturday in the park with my family.
09:35That sounds wonderful.
09:37Anna, you mentioned last week that you were celebrating your birthday.
09:40How was it?
09:41Oh, that was fantastic.
09:44I had dinner at my favorite restaurant with some friends.
09:47If you visit Sweden, you must try this restaurant.
09:51It was
09:52So, um a lot of things happening.
09:54Not only there's, you know, real-time translation.
09:56Whoops.
09:58How do I back to Okay.
10:00Um not only there's real-time translation that is happening, uh but we see that it's in a multi-speaker setting, you know, low latency, like right after the user starts speaking, the translation kicks off, so that it doesn't feel like it's really turn by turn and robotic and you're afraid to interrupt because,
10:15you know, once you start speaking, the model will start catching your translation directly.
10:19So, now I'll show some other applications where we see that we have one single speech-to-speech model for many conversational frontiers.
10:26So, two products that I want to highlight here are the same model that we power search live for everyday conversations in any language.
10:33It's the same model that we use to power developer experiences in our live API.
10:38As you'll see in the videos, both use cases are extremely different, but because we try and make this this model or this speech-to-speech model very promptable and universal, you can see how it adapts to the different use cases.
10:49So, let me just play the first video.
11:12So, to highlight in this video, many things are happening.
11:14First, the multimodal, you know, it's real-time uh video and audio streaming in, so that you don't have to actually describe what the sofa looks like.
11:22But then the multimodal out, you see that there was a tool called made to pull up the right search cards if the user wanted to know more information.
11:28And then finally, the ITN component, you know, we're localizing the speaker's Spanish to Spain Spanish.
11:34And also, when we do the kind of responding in Spanish, we're also smart enough to know that some words are better to be left in English, like the mid-century
11:42type of word.
11:43It didn't translate it directly to Spanish cuz it knew that the term was commonly known in English.
11:48Sorry, in Spanish.
11:49So, this is one example of how users are using search live for like info-seeking questions.
11:54Now, let me show a very different demo on how we do this for conversational voice agents.
12:02Oh.
12:04Sorry.
12:14Hi, I've blown a tire and had to pull over to the side of the road.
12:18Can you help?
12:20Hello, you've reached Auto Transform Insurance Roadside Assistance.
12:24My name is Ace.
12:26I understand this is stressful, but we'll take care of everything.
12:29To get started, could I please have your name and your policy number?
12:32Uh my name is Richard Jenkins.
12:34I I don't actually have my policy number to hand.
12:39No worries, Richard.
12:40I can look you up with your registration plate and postcode instead.
12:44Could you provide those for me, please?
12:46Whoa, that was a bit close.
12:48Um yeah, my registration plate is BD21 XYA and uh my postcode is SN48ZX.
13:03Policy details BD21 XYA.
13:07Thank you.
13:08For your safety, please stay clear of the road.
13:11I've found your details and I see you're in a blue Mini Cooper F-Series.
13:15Is that the vehicle you're in?
13:16Yeah, yeah, that that's the vehicle.
13:20Got it.
13:21Okay, so I'll pause it here, but as you can see, very different things are happening under the hood with the same model.
13:26You know, it's more about alphanumeric accuracy for complex kind of like postcode numbers or addresses.
13:32There's like this feature that we have called proactive audio, which is the idea that the LLM knows when or when not to respond to an external input.
13:40So, for example, if you're in a conversation and someone else is talking or there's background noise happening in the back, the model knows to not stop or cut itself short because there's external noise happening.
13:50Because we realized that the majority of these conversations are not happening in an office room with like no noise, you know, in a sealed environment.
13:57They're happening on the go, on the train, while you're on a walk.
14:00Um and that's the type of experiences that we want to facilitate with the speech-to-speech model.
14:05Um lastly, um another thing we're really excited about is not only voice out, but also multimodal out.
14:11We believe that, you know, the true AGI conversational frontier will also require visual presence.
14:16So, we were excited to launch our first kind of pilot demo with City um in Cloud Next, which supports customized real-time avatars.
14:24Um and you can kind of like personalize anything from a hyperrealistic human to a cartoon appearance and everything in the middle.
14:32And powered by the same speech-to-speech model we've been showing, it allows to have, you know, low-latency, multilingual lip-syncing, and a really kind of like uh ongoing fluid conversation that has visual presence.
14:42So, let me show you a demo that also brings our Venn diagram together and what we're excited about.
14:48All this is reminding me of my daughter's college fund.
14:51How are we tracking on that?
14:55You're tracking well, Jackson.
14:56I've also identified a new opportunity that may get you there even sooner.
15:01Oh, and I can see Lisa just joined you.
15:03Hi, Lisa.
15:04She must be so excited about her college acceptance.
15:07Congratulations.
15:08It's wonderful to see your savings goals coming to life.
15:11Excelente.
15:12Las cosas han estado tan inestables últimamente.
15:18Sí.
15:19Los servicios de tecnología y comunicaciones han mostrado un desempeño sólido en lo que va del año.
15:25That's good news.
15:26I always joke that the user's audio in Spanish is worse than the audio model speaking back, but um this is kind of just to show how like our Venn diagram of combining, you know, multimodality
15:36in and out, you know, tool calling to like pull up the relevant examples from the user, and also conversational fluidity with ITNN are starting slowly to come together in these types of demos that we're excited to keep pushing the frontier of.
15:49Um so with this parting thought, I guess last kind of thought that we have for you is that we believe that AGI will not be typed, that it will be spoken.
15:57Um and for it to be spoken, there's a lot of things that need to work together in a single promptable, versatile model that allows a user to switch between all the sorts of conversation modes that we're looking at, right?
16:08From translation to taking action to brainstorming to rambling, and we truly believe in the power of these speech-to-speech models to achieve that like seamless switching.
16:17Um so we're excited to push the frontier on that.
16:19So if you're excited or want to learn more, please come talk to us, and thank you so much for coming.
16:38[music]