IMO a couple of possibilities here. It comes down to Twitch having one of the best datasets in the world for real-time, multi-modal human interaction. A Twitch stream isn't just the video asset - it's live speech, live chat, live follows, subs, polls, etc.
Voice models are good, but they are largely trained on convos between one human and one AI. Training on community interactions vs. solo interactions presents a new frontier for voice models. This is highly valuable to Amazon since a human conversational behavior dataset can be used for other verticals like Alexa, Amazon Shopping, and Amazon Gaming.
In terms of Twitch itself - they will not create "AI streamers" as they fully understand streaming is about the live human persona. Instead, they might use this for product features like AI co-pilots for streamers (suggest jokes, auto-clip good moments, create polls at the right time, etc.). I see them already leaning into this via the Auto-Clip Alpha.
Twitch and Amazon are apparently planning to use streams to train Generative AI.
Multiple sources tell me it will be announced in the coming months, despite some internal friction against it. Creators may be able to opt out of some training data.