My son wanted me to make a "Verity" interactive Halloween display. Verity is a "helpful ai assistant" minecraft mod that turns into a monster.
So I decided to see if I could make a real-time AI chat bot using my local AI models.
I used NVIDIA Nemotron Speech Streaming English 0.6B for the speech to text
that sent the text to qwen3-4b (I used a smaller model so verity could respond in under a second, I wanted it to be as natural as possible)
then the response from qwen3-4b went to Qwen3-TTS-12Hz-0.6B-Base for text to speech.
qwen3-tts has a nice feature that also lets you do "deepfake" or voice cloning. This was great so I could get the exact verity voice that my son was familiar with.
overall I think it went pretty well. And based on the logs, kids loved playing with it.
Verity would answer any question the kids had (since the model's are not censored, I replaced all bad words between its "brain" and its "mouth" so the kids could try to trick it into saying bad words, but the words would be replaced with silly words like "pickle" or "shoe" before converted into audio.
Verity would start happy, but get more and more frustrated if you tell him he's wrong, or are mean to him. Eventually saying "You know what? take your candy and get out!"


No comments:
Post a Comment