Powering Speaking Characters with Generative AI — Google for Builders Weblog

Posted by Jay Ji, Senior Product Supervisor, Google PI; Christian Frueh, Software program Engineer, Google Analysis and Pedro Vergani, Employees Designer, Perception UX

A customizable AI-powered character template that demonstrates the ability of LLMs to create interactive experiences with depth

Google’s Accomplice Innovation crew has developed a collection of Generative AI templates to showcase how combining Giant Language Fashions with present Google APIs and applied sciences can remedy particular trade use circumstances.

Speaking Character is a customizable 3D avatar builder that enables builders to carry an animated character to life with Generative AI. Each builders and customers can configure the avatar’s character, backstory and information base, and thus create a specialised knowledgeable with a singular perspective on any given matter. Then, customers can work together with it in each textual content or verbal dialog.

An animated GIF of the templated character from the Talking Character demo. ‘Buddy’, a cartoonish Dog, is shown against a bright yellow background pulling multiple facial expressions showing moods like happiness, surprise, and ‘thinking face’, illustrating the expressive nature of the avatar as well as the smoothness of the animation.

As one instance, now we have outlined a base character mannequin, Buddy. He’s a pleasant canine that now we have given a backstory, character and information base such that customers can converse about typical canine life experiences. We additionally present an instance of how character and backstory could be modified to imagine the persona of a dependable insurance coverage agent – or the rest for that matter.

An animated GIF showing a simple step in the UX, where the user configures the knowledge base and backstory elements of the character.

Our code template is meant to serve two important objectives:

First, present builders and customers with a take a look at interface to experiment with the highly effective idea of immediate engineering for character improvement and leveraging particular datasets on prime of the PaLM API to create distinctive experiences.

Second, showcase how Generative AI interactions could be enhanced past easy textual content or chat-led experiences. By leveraging cloud providers equivalent to speech-to-text and text-to-speech, and machine studying fashions to animate the character, builders can create a vastly extra pure expertise for customers.

Potential use circumstances of the sort of expertise are various and embody utility equivalent to interactive artistic software in creating characters and narratives for gaming or storytelling; tech assist even for advanced techniques or processes; customer support tailor-made for particular services or products; for debate observe, language studying, or particular topic training; or just for bringing model belongings to life with a voice and the power to work together with.

Technical Implementation


We use a number of separate expertise elements to allow a 3D avatar to have a pure dialog with customers. First, we use Google’s speech-to-text service to transform speech inputs to textual content, which is then fed into the PaLM API. We then use text-to-speech to generate a human-sounding voice for the language mannequin’s response.

An image that shows the links between different screens in the Talking Character app. Highlighted is  a flow from the main character screen, to the settings screen, to a screen where the user can edit the settings.


To allow an interactive visible expertise, we created a ‘speaking’ 3D avatar that animates primarily based on the sample and intonation of the generated voice. Utilizing the MediaPipe framework, we leveraged a brand new audio-to-blendshapes machine studying mannequin for producing facial expressions and lip actions that synchronize to the voice sample.

Blendshapes are management parameters which might be used to animate 3D avatars utilizing a small set of weights. Our audio-to-blendshapes mannequin predicts these weights from speech enter in real-time, to drive the animated avatar. This mannequin is educated from ‘speaking head’ movies utilizing Tensorflow, the place we use 3D face monitoring to be taught a mapping from speech to facial blendshapes, as described on this paper.

As soon as the generated blendshape weights are obtained from the mannequin, we make use of them to morph the facial expressions and lip movement of the 3D avatar, utilizing the open supply JavaScript 3D library three.js.

Character Design

In crafting Buddy, our intent was to discover forming an emotional bond between customers and its wealthy backstory and distinct character. Our purpose was not simply to raise the extent of engagement, however to exhibit how a personality, for instance one imbued with humor, can form your interplay with it.

A content material author developed a charming backstory to floor this character. This backstory, together with its information base, is what offers depth to its character and brings it to life.

We additional sought to include recognizable non-verbal cues, like facial expressions, as indicators of the interplay’s development. As an example, when the character seems deep in thought, it is a signal that the mannequin is formulating its response.

Immediate Construction

Lastly, to make the avatar simply customizable with easy textual content inputs, we designed the immediate construction to have three components: character, backstory, and information base. We mix all three items to 1 massive immediate, and ship it to the PaLM API because the context.

A schematic overview of the prompt structure for the experience.

Partnerships and Use Instances

ZEPETO, beloved by Gen Z, is an avatar-centric social universe the place customers can totally customise their digital personas, discover vogue tendencies, and interact in vibrant self-expression and digital interplay. Our Speaking Character template permits customers to create their very own avatars, gown them up in numerous garments and equipment, and work together with different customers in digital worlds. We’re working with ZEPETO and have examined their metaverse avatar with over 50 blendshapes with nice outcomes.

A schematic overview of the prompt structure for the experience.


“Seeing an AI character come to life as a ZEPETO avatar and converse with such fluidity and depth is really inspiring. We consider a mixture of superior language fashions and avatars will infinitely broaden what is feasible within the metaverse, and we’re excited to be part of it.”– Daewook Kim, CEO, ZEPETO


The demo isn’t restricted to metaverse use circumstances, although. The demo reveals how characters can carry textual content corpus or information bases to life in any area.

For instance in gaming, LLM powered NPCs might enrich the universe of a sport and deepen person expertise by pure language conversations discussing the sport’s world, historical past and characters.

In training, characters could be created to characterize completely different topics a scholar is to review, or have completely different characters representing completely different ranges of problem in an interactive academic quiz state of affairs, or representing particular characters and occasions from historical past to assist folks study completely different cultures, locations, folks and instances.

In commerce, the Speaking Character package might be used to carry manufacturers and shops to life, or to energy retailers in an eCommerce market and democratize instruments to make their shops extra participating and customized to present higher person expertise. It might be used to create avatars for purchasers as they discover a retail atmosphere and gamify the expertise of buying in the true world.

Much more broadly, any model, services or products can use this demo to carry a speaking agent to life that may work together with customers primarily based on any information set of tone of voice, appearing as a model ambassador, customer support consultant, or gross sales assistant.

Open Supply and Developer Assist

Google’s Accomplice Innovation crew has developed a collection of Generative AI Templates showcasing the chances when combining LLMs with present Google APIs and applied sciences to resolve particular trade use circumstances. Every template was launched at I/O in Could this 12 months, and open-sourced for builders and companions to construct upon.

We are going to work intently with a number of companions on an EAP that enables us to co-develop and launch particular options and experiences primarily based on these templates, as and when the API is launched in every respective market (APAC timings TBC). Speaking Agent can even be open sourced so builders and startups can construct on prime of the experiences now we have created. Google’s Accomplice Innovation crew will proceed to construct options and instruments in partnership with native markets to broaden on the R&D already underway. View the undertaking on GitHub right here.


We want to acknowledge the invaluable contributions of the next folks to this undertaking: Mattias Breitholtz, Yinuo Wang, Vivek Kwatra, Tyler Mullen, Chuo-Ling Chang, Boon Panichprecha, Lek Pongsakorntorn, Zeno Chullamonthon, Yiyao Zhang, Qiming Zheng, Joyce Li, Xiao Di, Heejun Kim, Jonghyun Lee, Hyeonjun Jo, Jihwan Im, Ajin Ko, Amy Kim, Dream Choi, Yoomi Choi, KC Chung, Edwina Priest, Joe Fry, Bryan Tanaka, Sisi Jin, Agata Dondzik, Miguel de Andres-Clavera.

Leave a Reply

Your email address will not be published. Required fields are marked *