Higgs Audio v3: radio TTS workflow and Gemini comparison

Jeff the Radio DJ Dude and Jeff Hansen discuss Higgs Audio version 3 for local AI voice tracks. They play original voice samples and cloned output, then compare the results with ElevenLabs and Gemini. The host considers Higgs the closest local TTS option he has heard for radio, but says quality remains inconsistent.

Their workflow uses Gemini to analyze a reference recording's cadence, tone and sentence length, then produce writing rules for that voice. Scripts based on those rules feed the cloned voice in Higgs. The speakers report better delivery than they achieved with scripts originally written for Gemini voices. This workflow still uses Gemini for audio analysis, even though speech generation runs on Hansen's hardware.

They describe inline emotion and delivery controls, with one emotion tag per script in their setup. Hansen reports an output window of about 30 seconds. He is testing chunking for longer scripts, but says joins can introduce audible differences. The host also warns about stray audio and recommends checking generated tracks before broadcast.

Hansen offers a Linux installation guide for viewers with suitable GPU hardware. The video does not specify a minimum GPU requirement, and the host warns that generation may be slow on viewers' systems. They also announce an experimental trial for 10 RoboDJ v3 users through an API to Hansen's server; running that integration on users' own hardware is described as a future goal.