Paige Bailey’s Tour of AI Video, Audio and App Creation

Hannah Fry welcomes Google DeepMind developer-relations lead Paige Bailey for a hands-on tour of video, audio and app-building tools. The demonstrations describe products as they existed at the time of recording.

They revisit an earlier Veo demonstration and compare it with Veo 3 using the same prompt for a racing scene. Fry notices more detailed buildings and lighting; Bailey highlights generated sound.

Bailey says earlier generated video could show inconsistent physics or light. A more detailed prompt can help guide a result, and she demonstrates using a model to rewrite a short prompt with additional description.

At the time of the demonstration, Bailey says the publicly available Veo clips were around eight seconds. She describes testing a style in a short clip before assembling a longer sequence.

In the racing result, the speakers point to reflections, moving light and sounds such as tires and sirens. Their observations concern the example shown; they do not establish that generated physics or audio are always correct.

They try a more difficult spaghetti-eating scene with several people. Bailey says an image can also serve as an initial frame, and an image model can modify that frame before video generation.

Bailey attributes improvements partly to combining video, sound and descriptions in training. Fry describes the value of relating different modalities, such as a tree’s image, words and surrounding sounds.

They watch the generated contest and inspect faces, hair, motion and sound. Fry is particularly interested in whether an object remains consistent across frames; Bailey describes the model as attempting to maintain that consistency.

Bailey mentions synthetic data from simulated environments as one possible training source. She also explains that matching audio and video involves multimodal information rather than treating every output in isolation.

They move to Flow, a filmmaking interface. Bailey demonstrates text-to-video and frame-based workflows, clip assembly and camera controls; she describes it as a more specialized environment than the general Gemini app.

Fry asks about deepfakes, scams and fabricated news events. Bailey describes filters, restrictions and watermarking used in the demonstrated tools. Such measures are safeguards, not proof that every misleading output will be prevented.

Bailey introduces text-to-speech with controllable voice, language and style. They compare it with an older WaveNet example made using Fry’s voice and discuss how the systems differ.

She says training data from earlier speech work can remain useful even when model architectures change. Data can include text, audio and descriptions of tone or delivery.

In AI Studio, the pair generate the same sentence in friendly, romantic, angry and grieving styles, then try a French response. They note that one result does not fully match the requested emotion and could be refined.

Bailey says the interface can show code for reproducing an experiment through an SDK. Availability, pricing and product controls may change after the recording.

They demonstrate a live assistant with screen sharing and search. It describes a video on screen, then uses search to answer a question about creative tools; the example shows a workflow, not a guarantee that every description is accurate.

Fry tries camera-based interaction and mentions using it to look at houseplants. Bailey demonstrates sharing a notebook and asking the assistant to explain visible code.

Bailey then shows an app-building workflow. The pair describe a fictional spaghetti contest website with generated images and a sign-up form; the tool generates a project and a preview.

The generated page uses a playful style and fictional contest highlights. Fry and Bailey discuss refining the result and, in the demonstrated environment, deploying it to Cloud Run. Generated content and code still require review for accuracy and security.

Bailey argues that automation could let developers spend more time on ideas and product experience, while assisting with maintenance. She expects creative tools to help more people express ideas across disciplines.

Fry closes by reflecting on the experience of seeing audio, video and language tools together. She is excited to experiment with ideas she previously lacked time to build.