It seems like things are going this way. I sure could use this to automatically create drum tracks for my instrumental tracks. No matter how hard I try I can't seem to create decent drum tracks. I'm aware of most of the drum apps out there, but if there's an easier way I'll take it. Moises AI Studio looks and sounds good to me. One major drawback ls the price.

 

I was looking at the Pro price and it was $16 a month, but required a yearly which was $199 a yr. If you pay monthly, it's $29.99. Also seeing the small print, it is a hassle to stop automatic payments.

 

I marked your comment "i love it" because at least it is honest, that you would take the easy route if you could.

It explains why many people would try to do this too. What it really points out though, is a need for an AI detection method so people stop wasting their money and credibility trying to get away with it, and other peoples' time to remind them they aren't.

 

I just demoed moises, and the few passes i did with the drums sounded bad. Grainy, no punch, AI sounding stuff.

 

It’s not really enjoyable to make music now. It takes a lot of time, it takes a lot of practice, you need to get really good at an instrument or really good at a piece of production software. I think the majority of people don’t enjoy the majority of the time they spend making music.

 

I've played with this a little. Could not really generate anything usable.

 

Was there a time when it didn't require time and practice? If you don't enjoy it at all, don't do it. A good production (be it a mix that's better than a "demo level" or a song that your band performs good enough) takes time and of course practice. That's why people who are interested in music not just as a listener, but also as a producer/musician, learn, gain experience. For most people, it's a cool hobby, but it's not something that brings instant gratification. There are many other hobbies/activities that can provide that.

 

He's obviously talking about people who are "making music" using prompts to AI generate stuff.

 

Save the money you spend on the AI; after a year of saving, you can buy a proper drum machine, software, or hardware. If you have weaknesses with drumming, it's certainly helpful to acquire knowledge and learn. The more you leave it to the machines, the more they become less independent. For decades, we've managed without AI and still produced a considerable amount of music.

 

How is it possible to watermark MIDI ? Except for sysex, MIDI is a strict norm from the 1980s without any empty bytes or room for additional data. And even if possible, how could this data transmitted to the sound generating device ?

 

Unfortunately, my expertise is not sufficient to assess the AI results; never underestimate the technological progress in copy protection and copyright.

Spoiler: Source: Grok KI

Watermarking MIDI data, such as in Standard MIDI Files (SMF), is indeed challenging due to the protocol's rigid structure from the 1980s, which lacks dedicated space for arbitrary metadata beyond System Exclusive (SysEx) messages. However, researchers have developed techniques that embed watermarks (e.g., for copyright protection or data hiding) by making imperceptible modifications to existing MIDI event parameters, without inserting new bytes or events. These methods exploit the fact that certain fields, like note velocities or timings, have some "redundancy" in human perception—small changes don't noticeably alter the music. Below, I'll outline key approaches based on established research, focusing on those that avoid SysEx.

Techniques for Embedding Watermarks
These methods typically involve steganography or data hiding, where the watermark (e.g., a binary string representing a logo, ID, or message) is encoded into the MIDI data. The embedding is designed to be robust against minor edits, reversible in some cases (allowing exact recovery of the original), and imperceptible to listeners.

1. Velocity Modulation:
   • How it works: Note velocities (ranging from 0 to 127, representing volume or intensity) are slightly adjusted to encode bits. For example, in one approach, velocities within musical measures are compared to reference values based on beat types (e.g., strong beats have higher velocities). The difference between a note's velocity and its reference allows embedding 1–3 bits per eligible note by shifting the velocity closer to or farther from the reference, while staying within the original range to avoid audible changes. A leading bit helps detect embedded data during extraction.
   • Capacity and imperceptibility: This can embed hundreds to thousands of bits per file (e.g., 415–5,994 bits in tested 4–30 KB files), with an average of 1.7–3.2 bits per embeddable note. Changes are confined to natural dynamic variations, so the music sounds unchanged. File size remains the same.
   • Examples from research:
     • One method uses least significant bits (LSBs) of velocities or pairs notes to add/subtract small values (e.g., +1 to one note's velocity and -1 to another's), balancing the overall volume.
     • Another refines this by measure-based references (e.g., strong/weak beats in 4/4 time), ensuring adjustments align with musical structure.
2. Duration or Delta Time Modulation:
   • How it works: Delta times (variable-length values indicating ticks between events) or note durations are subtly altered. One reversible technique shifts the binary representations of non-zero delta times and the file's division value (ticks per quarter note) left by a fixed number of bits (e.g., 1–3), freeing up lower bits for watermark data. Bits are then inserted into these slots, often using pairs of adjacent delta times. After embedding, the file structure stays intact.
   • Capacity and imperceptibility: Embedding capacity scales with file complexity (more events mean more delta times), and shifts are limited to tolerances that cause negligible timing changes (e.g., fractions of a beat). The original can be restored by right-shifting during extraction. No file size increase occurs.
   • Examples from research:
     • Pair consecutive notes and adjust durations (e.g., add a small tick to one and subtract from the next) to encode a bit without changing total length.
     • Shift-based methods ensure blind reversibility and minimal audible impact, passing perceptual tests like χ² analysis.
3. Other Parameter-Based Methods:
   • Some combine velocity and duration for hybrid robustness, or use list-based steganography (e.g., reordering non-essential event lists subtly to encode data via permutations). These maintain the strict MIDI norm by working within existing bytes.

Robustness varies: These techniques resist basic attacks like tempo changes or velocity normalization but may fail under heavy editing. They outperform SysEx-based methods in stealth, as SysEx can be easily stripped.

Read the terms and conditions before using anything; they're obviously difficult to understand, or ask here in the forum before you get into trouble:

8.3 Moises Generated Content.

Subject your compliance with these Terms, Moises grants you a non-exclusive, non-transferable, perpetual right and license to use the Moises Generated Content solely in combination with other lyrics and sounds for songwriting and music production purposes to create new lyrics, musical works, sound recordings, and audiovisual works. This means that you may modify, reproduce, publicly perform, distribute, transmit, communicate to the public, create derivative works of, and otherwise use such Moises Generated Content, including for commercial purposes.

However you may not (a) use or sublicense the Moises Generated Content in isolation as suggestions, lyrics, sound effects, audio, or as source material for any other form of sample (even if you modify the Moises Generated Content); (b) use Moises Generated Content in a manner competitive to Moises or its licensors; (c) use Moises Generated Content or Output to train or tune artificial intelligence or machine learning models either as the sole input or as part of a larger dataset;

or (d) sublicense, sell, loan, share, lend, broadcast, rent, lease, assign, distribute, or transfer all of the Moises Generated Content to a third party except as incorporated into a new composition. For clarity, “Moises Generated Content” does not include the Moises Collection.

Source and read more about the Terms and Conditions: https://help.moises.ai/hc/en-us/articles/7401394754962-Terms-Conditions

 

Thanks @PulseWave.
In my opinion, this MIDI watermarking should be very easy to defeat but, if someone has the -far from stellar- competence level needed to do so, they will not use that kind of service to begin with.

 

Of course midi can have hidden id's. It takes 2 seconds to search and find out the truth. Whoever is feeding you all this bullshit saying it cant, is not someone you should ever trust again. Too many salespeople around these ai threads here smh

If you think you are the smartest guy in the room who's going to be sneaky and erase the watermark's, id's and copyright, without totally destroying the audio and midi, and then lie to secretly fool everyone into believing your music is 100% original and ai-free, you're the delusional fool and begging for criminal and civil lawsuits, along with blacklisting yourself and ending all your chances of ever having a career of any capacity outside of janitorial in the entertainment industry.

The only way this ai works for the consumer is if you don't commercially distribute it and you are totally open, upfront and honest about it being ai generated or "assisted" music lol. But then that means even if it isn't removed by a website or banned in a country or something, you still will never have rights to collect any royalties or accolades no matter how many sales or streams you get. Imagine getting 10 million streams and being paid NOTHING! So now you work for those streaming websites, generating money for them but not yourself. Sounds like a great plan!! lol

As a rule of thumb, a good friend said it best DON'T FEED THE AI CREDITS COOKIE MONSTER!

 

ai bad

 

on the serious matter moises should be buried in the depths of the internet and never return for the 2020 vocal sep "quality"

 

That was a quote from the suno ceo haha

 

ion wanna get into the flames, but i've sucessfully used it to split vocals from a AudioSex fellow member's finished tune and got pretty decent results, worth to mention, it wasn't a clean pristine vocal take i was after by the way, that remix was already gritty.

 

I got you. (Another time please add that you are quoting, or at least use quotation marks.)

 

Well hell, guess I'm going to have to keep plugging away with existing drum software for now. I've been out of music production for many years, and have decided to get back into it as a hobby in my old age. Years ago I bought a drum software app that would create a drum track automatically to accompany an existing instrument track. Unfortunately, I lost it and can't find it anywhere, and I can't even remember the name of it. I remember it worked really well, and was the first of it's kind. I also remember the developers were working on a bass version, but that's all I can remember. If anyone can help me remember what it was I'd appreciate it. Thanks.