**Dissertation now available** 📜: 📽️: ⏯️: It's a 332-page tome, but I have summarized it in this thread 👇 1/n

@shinjiw_at_cmu @chenwanch1 Congratulations! Very comprehensive paper and well deserved recognition :)

@giffmana Diwali bonanza

@lorenlugosch Nice! 💪

@huckiyang @ieeeICASSP Thanks for confirming! I was just confused because my submissions on CMT showed "under review"

@_josh_meyer_ I tried a few of these myself, and I must say Google has really nailed this one. One thing I noticed, however, is that the "male voice" is always the expert in the generated podcasts (or at least in the ones I have seen so far). Would be nice if this could be randomized!

@janekm @awnihannun @kyutai_labs @dongzha35524835 Thanks for the pointer; I was not aware of this work. Yeah, small POCs are easy. But as these folks found: "as we scale to bigger datasets or models, scaling laws may imply greater loss or degeneration." This is the harder problem to solve I think.

@janekm @awnihannun @kyutai_labs I can only comment on the research :) Even academic labs have shown POCs for the idea (e.g., SpeechGPT by @dongzha35524835). It was well known that speech instruct tuning works with discrete tokens. OAI probably also invested in RLHF with speech, which IMO is the next frontier.

RT @Ahmad_Al_Dahle: 📣 You can now have a conversation with Meta AI using voice. It’s super fast, connected to the web, natural and conversa…

RT @ArxivSound: ``M-BEST-RQ: A Multi-Channel Speech Foundation Model for Smart Glasses,'' Yufeng Yang, Desh Raj, Ju Lin, Niko Moritz, Junte…

@girlknowstech Xitter is dead. Only bots roam this space now.

@JonathanLeRoux @ieeeICASSP @IEEE_AASP ICASSP subject areas always baffle me. Why is there only 1 blanket area for "ASR" and why are there no new areas for LLMs? It seems these haven't been updated in 10+ years!

RT @ieeeICASSP: The #ICASSP2025 paper submission deadline is 9 September 2024. No new submissions will be accepted after this deadline. How…

@jxmnop The compute actually happens on super-fast SRAM, so the actual movement of tensors is between the HBM and the SRAM. Unfortunately the SRAM is small and fusing operations is complicated in general, as others have pointed out.

@JonathanLeRoux @ieeeICASSP @IEEEsps Oh this is news to me! Need to update all the overleaf projects now 😅

Every once in a while I find a thread on this website which makes all the spam bearable. Couldn't help sharing!

@ziqiao_ma I think it's more about ignorance than privilege.