-10
How can I change the voice of a person in a recorded audio clip with local-only AI? (programming.dev)
submitted 1 month ago* (last edited 1 month ago) by to c/Aii@programming.dev
5 comments fedilink hide all child comments
 

Goal: Take a recording of someone speaking, pass it through a CLI or GUI, output a recording with the speaker's voice replaced by an AI voice.

I know there are closed-source or hosted services that do this, but they are a privacy risk and most surely also continue training their models on what users submit. The opensource solutions I have seen (last I checked 6 months ago or so) were all extremely outdated (around the beginning of the ChatGPT release) and required training using thousands of voice samples.

Surely somebody has done this before and didn't need to be an expert in the field to do so...

Non-solutions:

  • hosted services (elevenlabs, chatGPT, Gemini, ...)
  • "train your own AI"

all 5 comments

sorted by: hot top controversial new old
[–] 3 points 1 month ago

combine whisper speech-to-text with any one of the numerous local text-to-speech solutions like picotts, piper, kitten, pocket-tts, outetts etc.

  • permalink
  • fedilink
  • source
    •
    [–] 1 point 1 month ago (1 child)

    What's wrong with the old school approach of adjusting pitch & such?

  • permalink
  • fedilink
  • source
  • hideshow 2 child comments
    •
  • [–] [S] -1 points 1 month ago (1 child)

    That doesn't give a new voice and I can't send my friends messages sounding like a specific personality

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 2 points 1 month ago (1 child)

    I assumed you want to make person unidentifiable, in which case tone and pitch changes will not erase enough of biometrics.

    If you want to make it sound like someone specific, then it's exactly what you said in the last line, you need to train your own spoofing models. Also you sound like you want to do something I can't endorse, so it's good that you don't want to train a model

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] [S] 1 point 1 month ago

    Also you sound like you want to do something I can't endorse, so it's good that you don't want to train a model

    Not sure what that means. I want to send a message sounding like Trump or respond with a joke in the voice of Ghandi to friends.

    Anonymizing would also be a usecase. Then I could replace the audio of videos I want to make after narrating them, record an anonymous podcast, or send voice messages into groups I'm a part of with a different voice.

  • permalink
  • fedilink
  • source
  • parent