zaptrem

21 ms·

1. 271.▲
   by zaptrem 3y ago

   If the future of music was truly just typing some text into a box and taking or leaving what the machine gives you that would be kinda depressing. We want you to be able to upload recordings of your real instruments and do all sorts of cool
2. 272.▲
   by zaptrem 3y ago

   Neat! Is this a diffusion model, language model (e.g., SoundStorm from your former colleagues), or something else?
3. 273.▲
   by zaptrem 3y ago

   Thanks! We have a BPM assist that can enforce rhythm as well, so you could try that, too!
4. 274.▲
   by zaptrem 3y ago

   Which providers would you prefer? We tried Twitter last night but it wasn't working for some reason (kept redirecting immediately with no oauth page).
5. 275.▲
   by zaptrem 3y ago

   This can definitely be done. There are approaches that turn the decoder part of the autoencoder into another diffusion model. The drawback is that's much more expensive computationally. We think there's still a lot of room for bet
6. 276.▲
   by zaptrem 3y ago

   This is exactly what makes it so exciting for us!
7. 277.▲
   by zaptrem 3y ago

   If this were the future that would be kinda depressing. I think the best, truly catchy songs and those that truly connect with people will continue having a significant human element. I see this as similar to the invention of Photoshop exce
8. 278.▲
   by zaptrem 3y ago

   Upload an instrumental track, select it, then click "Use as Rhythm Control." Once you do that, you can give the model any new prompt and it should use the same rhythm (you may need to adjust the control strength depending on genre
9. 279.▲
   by zaptrem 3y ago

   This would be so cool, but we need to think more about how we could do it and make enough money in the future to train more models with even cooler features.
10. 280.▲
    by zaptrem 3y ago

    Thanks! We have lots of fun dials for people who want them but they're all hidden by default and shouldn't be needed.
11. 281.▲
    by zaptrem 3y ago

    Rhythm Control can do this for a drum line, and we have a variations feature that should be able to do this for instruments as well.
12. 282.▲
    by zaptrem 3y ago

    The audio is 44.1khz stereo, but all of us use autoencoders so the songs will fit in a transformer's context window, and huge compression will affect quality. We're definitely working on better ones, though!
13. 283.▲
    by zaptrem 3y ago

    Weird. We pushed a BPM assist feature last night that may have unforeseen consequences for genres we didn't test (we tried pop, edm, classic rock). I'll turn it off by default for now. Try checking the instrumental box too.
14. 284.▲
    Show HN: Sonauto – A more controllable AI music creator(sonauto.ai)

    454 points by zaptrem 3y ago|235 comments
15. 285.▲
    by zaptrem 3y ago

    This is quite similar to the models used by all the AI music generators. Some feed the tokens into a language model to generate music, some replace the tokenization part with an alternative that gives a continuous representation for diffusi
16. 286.▲
    by zaptrem 3y ago

    They thought of this! https://en.wikipedia.org/wiki/Tuned_mass_damper
17. 287.▲
    by zaptrem 3y ago

    This is Xiaomi we're talking about. They've got #2 covered. The faux self driving has gotten pretty good lately. I tried routing to random places in downtown SF yesterday for an hour while on a call and for the first time there we
18. 288.▲
    by zaptrem 3y ago

    "Lei claimed the standard version of the SU7 beat the Model 3 on more than 90% of its specifications, except on two aspects that he said it might take Xiaomi at least three to five years to catch up with Tesla on." The article doe
19. 289.▲
    by zaptrem 3y ago

    Not my area of expertise but I'd assume it becomes a decision tree or something. Edit: lol https://news.ycombinator.com/item?id=39868508
20. 290.▲
    by zaptrem 3y ago

    Transformer LLMs are just a bunch of MLPs (linear layers) where you sometimes multiply/softmax the output in a funny way (attention). In other words, they're arguably more "vanilla deep net" than most architectures (e.g.
21. 291.▲
    by zaptrem 3y ago

    Idk training a model like this is kinda trivial for most shops. Afaik there are open source ones that do the same thing. We trained one with similar audio quality to this by accident while working on music generation.
22. 292.▲
    by zaptrem 3y ago

    I thought everyone (11Labs, open source, OpenAI even) already has human-level TTS models? Is there still an open challenge somewhere (e.g., is there a use case where a better model would make any difference?).
23. 293.▲
    by zaptrem 3y ago

    Agreed. The above is essentially a tech demo while we work on the more exciting/fulfilling uses we're planning for these models.
24. 294.▲
    by zaptrem 3y ago

    I mean the user wrote the lyrics (or maybe ChatGPT lol). People of course write more interesting ones (Just found a Ukrainian about Kyiv: https://sonauto.ai/songs/xRDqe57ZgT6QrFXiIzF0 ). There was another one about Fran
25. 295.▲
    by zaptrem 3y ago

    yeah, we're training a diffusion model for this here: https://sonauto.ai/ example 70s dolly parton song made by a rando earlier today: https://sonauto.ai/songs/t4XGcHqsYwS6B73MjtV5
26. 296.▲
    by zaptrem 3y ago

    try this one my friend and I are training: https://sonauto.ai/ audio quality is still WIP and UI is ugly but prompting should allow more detail. e.g., 70s Dolly Parton-style song rando made earlier today: https:/&#x2F
27. 297.▲
    by zaptrem 3y ago

    Sometimes that takes many hundreds of years (see Russia)
28. 298.▲
    by zaptrem 3y ago

    I think rule of thumb is 1GB VRAM per 1 billion params quantized to FP8.
29. 299.▲
    by zaptrem 3y ago

    But can I use it just by going `model.to("cerebras")`?
30. 300.▲
    by zaptrem 3y ago

    Hits take time to become hits.