AI & MLJavascriptWeb Development

Recreating a Voice Controlled Photo Analyzer in the Browser

The photo-analysis interface in Blade Runner is a memorable idea: move around an image by speaking to the computer and inspect details as the view changes.

My browser photo-analyzer project borrows that interaction idea and combines it with a real, bounded vision capability. An HTML canvas viewer provides navigation, the Web Speech API handles spoken commands, and TensorFlow.js runs pretrained COCO-SSD object detection.

Separate the interface from the model

The canvas controls determine how the photograph is displayed. Spoken navigation gives the viewer another way to receive commands. Neither operation needs to invent image content.

The object detector performs a separate task: identify likely objects and their locations using a pretrained model. Its output can then be presented in the same interface.

That division makes the application easier to understand. The voice interface controls the view; the model supplies detections; conventional browser code connects them.

Browser AI is still an application workflow

Running a model through TensorFlow.js puts inference inside the browser application. The application still needs to load the model, prepare the image, and handle the result alongside user interaction.

The speech component also deserves its own evaluation. A viewer command needs to be recognized and mapped to an intended action. Browser support and the speech service's behavior are separate concerns from object-detection accuracy.

I would not describe the entire experience as offline merely because the detector runs in the browser.

The cinematic effect has a limit

Zooming can make existing pixels easier to inspect. It does not recover details the original photo never captured. Likewise, a detected label is a model prediction rather than a guarantee about an object.

The project's “enhance” atmosphere should therefore remain an interface reference, without implying generative reconstruction or super-resolution.

Where it fits in the history

The retained repository history begins in October 2025. This project brings a pretrained vision model and a speech interface into a small web application, rather than training a new detector.

A public walkthrough would show the command flow and a rights-cleared example photograph with its detections. That would demonstrate the interaction without relying on film frames or private uploaded images.

Code and background