Soonbuilt

Maina

Copy a voice from half a minute of talking. It reads anything back, in English, Hindi or Urdu.

Maina: two voice engines, Natural and Light; step one, record or upload a voice; step two, pick a voice and type what it should say; step three, your clips, each watermarked as AI-made.

What it is

Maina copies a voice from a 20 to 30 second recording and reads any text back in it, in English, Hindi or Urdu. It scores the recording out of 100 before you save it, so you know how close the copy will be. Everything runs on your own Mac or PC, so the recording never leaves it, and on an everyday 8 GB laptop it switches to a lighter English voice. Every clip carries an inaudible watermark marking it as AI-made. Free.

What was hard

Consent you can check
Maina asks you to confirm a voice is yours, or that its owner said yes, before it saves it. In TallyReel Studio the speaker reads a consent sentence with a one-time code out loud, and the app checks both the words and that it is the same voice.
Three languages, one voice
The same cloned voice reads English, Hindi and Urdu, so one narrator can carry a channel in all three.
Marked as AI-made
Every clip carries an inaudible watermark, so a cloned voice can always be identified as synthetic.

Hear it

The recording Nine seconds of Emma, one of Maina’s built-in sample voices, the sample it learned from. · 9s
The copy, in English Maina reading a sentence Emma never recorded. · 8s
In Urdu Alpha, another sample voice, reading Urdu typed in Urdu script. · 8s

Built with

  • Python
  • VoxCPM2 on MLX and PyTorch
  • Perth watermark
  • Native window (pywebview)

Have an idea?
Let's see if it's worth building.

Find ideasBook a call