PEFT and contextual biaising for TTS domain adaptation
I'm working on a cascade based TTS model for a very narrow domain with specific noise environment and heavy jargon library. I've tried every providers ASR/TTS models and honestly it's not good. Really not good (high WER and tuti cunti). So as a ML engineer I'm trying to fine-tune TTS models (currently on QWEN3 ASR), but I'm a text guy, not an audio guy and so is my team haha. So we are learning and trying to build one cause on-the-shelf are bad but it's very difficult and we need to ship soon. Someone in that situation that as counsels or can direct me to GitHub discussions/Discord groups/events.... that may help me ? (I'm US based but I can travel anywhere if there is a real value and a real community of people blocked as us trying to look for a solution