Blog/second meeting - Githubissues

ludekcizinsky commented 1 year ago

Blog post about the second meeting and MusicLM paper.

vercel[bot] commented 1 year ago

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated
webpage	✅ Ready (Inspect)	Visit Preview	💬 Add your feedback	Feb 14, 2023 at 11:06AM (UTC)

ludekcizinsky commented 1 year ago

Mika: Ludek just reviewing the blog post, couple notes/ questions:

did you put a filter on the image haha?
regarding the text: let's try to keep the tags topic related (generative-ai, text-to-audio, etc. instead of second-meeting or aitu), otherwise it doesn't really make sense to use them for filtering. genearlly, always think which keywords are not already mentioned in the title, description, ... and other fields that one can search by. the tags should add value for the search in the sense that a viewer who is only interested in papers that we read about generative ai should easily find them
regarding training explanation: you explain audio lm and mulan, but the visualisation uses mulan, sound stream and bert - this might be confusing for readers. any particular reasons for this? in the summary i wrote on notion i have a quite high-level explanation of the role of each of the three models that we could use. but i wanted to hear your thoughts first

RE:

yes I did, do not like it? :D
yeah makes sense
Yeah I see what you mean. I wanted to stay as high as possible therefore when explaining AudioLm I did not explicitly say that the semantic tokens are extracte using W2v-Bert (btw. this is extension of the original Bert model for essentially encoding audio), and the accoustic ones using SoundStream. I could also perhaps explain more the role of Soundstream. Overall, I tried to put more emphasis on the types of tokens used and what information they encode rather than the name of the technology/paper that extracts them. In summary, in my eyes to solve this issue, we just mention explicitly in the explanation of AudioLM that extraction of either of the tokens is done by the given paper (W2v-Bert and SoundStream)

ludekcizinsky commented 1 year ago

another reason why I explained just audioLM and MuLan is because of the paper overview figure where you can see the two direct connections are AudioLM and MuLan

aitu-dk / webpage