AI Speech, Image & Video App

Vision Voice AI

One app that turns text into speech, images and short videos.

Vision Voice AI app screens
Built with
  • Flutter
  • Dart
  • flutter_tts
  • Pollinations API
  • Replicate API
  • video_player
  • gal
Platforms
Android, iOS
Year
2026

What it does

  1. Text-to-speech in English, Urdu, Arabic, French, German and Spanish, with adjustable rate, pitch and voice.

  2. Image generation through Pollinations.ai with five models (FLUX, FLUX Realism, FLUX Anime, FLUX 3D and Turbo), style presets and four aspect ratios.

  3. Text-to-video through Replicate's MiniMax Video-01 model, with an optional prompt optimizer and a built-in player.

  4. Generated images and videos save straight to the gallery.

How it’s built

The app has three tabs, one per tool, inside a single TabBarView. Each tool is its own screen with local state (setState), because the tools don't share data: a prompt typed for an image has nothing to do with the text being read aloud. Requests go out with the http package, straight from the app to each service.

Speech runs on the phone. flutter_tts drives the device's own text-to-speech engine: it works offline and costs nothing per request. Before every playback the screen sets the language, rate and pitch, it lists the voices installed on the phone, and a completion handler resets the controls when speaking ends.

Images come from Pollinations.ai in a single GET request. The prompt, the chosen style, the model, the width and height and a random seed are encoded into the URL, and the response body is the image itself. It is shown with Image.memory and saved to the gallery with the gal package after a permission check.

Video uses Replicate's prediction API: one POST creates a MiniMax Video-01 prediction, then the app checks on it every five seconds, downloads the finished clip to temporary storage and plays it with video_player.

Hardest part

Video generation doesn’t fit a normal request: a clip takes minutes, and the first response can contain either the finished video or only an ID to check later. The screen handles both. It asks Replicate to wait before answering, accepts 200, 201 and 202 as success and uses the output at once if it’s already there. Otherwise it polls the prediction until it succeeds, fails, is cancelled or reaches a ten-minute deadline, showing the seconds elapsed while you wait.

Replicate can return the output as a single URL or as a list, and a small helper reads both before downloading. Image generation has a 120-second timeout, a readable error message when a request fails, and the negative prompt added to the prompt text itself rather than sent as a separate field.

Next projectToneCraftRingtones & Wallpapers App