The topic of “Artificial Intelligence” (AI) was discussed as part of a lecture. During the follow-up, I came across, among other things, the website “Soundraw.io“, which claims to generate pieces of music with the help of AI on the basis of a parameterization specified by the user. But is it really “Artificial Intelligence” or just an expensive fake?
If it sounds too good to be true …
It almost sounds too good to be true: High-quality, ready-produced music at the push of a button for a small price. Soundraw promises that you only have to enter genre, instruments, mood, etc. and an AI will produce music on this basis – with a click on the menu item “Create Music” you can see for yourself right away.
The first impression is a “wow” – in a matter of seconds, several pieces of music are created, which definitely have mainstream potential and sound absolutely professionally produced. With one click, a lossless file in WAV format is created, which can then be used for videos, presentations etc.
It looks like a rather small audio loop library….
However, if you look and listen a bit closer, you will quickly notice that many music pieces sound surprisingly similar – is that a coincidence or is the provider perhaps “cheating” here? In any case, my curiosity was aroused – it reminded me all too much of a modular system, in which new combinations are constantly thrown together from prefabricated sound snippets on a metadata basis.
When analyzing the website I was more or less confirmed in my assumption. Right when checking the files retrieved from the server, it is noticeable that very many files are retrieved in M4A audio format. Furthermore, there are some metafiles with the file name “composition”, which describe respective concatenation of the loaded audio files. In effect, a combination of a manageable number of audio loops is “diced” here, which is then combined into a stereo mix. The result is professional sounding music.
The appearance allows the assumption that there is not really an AI at work here, but “only” a more or less clever algorithm. The low variance and high similarity of the music pieces speaks against this. It speaks much more for the fact that there is a basic set of ready-made “compositions” or that these are more or less put together on a random basis and if necessary only a fade is added to the end.
If you look a bit closer, you’ll find a reference to an Amazon S3 bucket – and if you follow that reference, you’ll find an XML file there that lists all the available sound snippets. There are pretty much exactly 1,000 of them from different genres. The file names describe music style, instrumentation etc.. Considering that there are currently about 11 different music genres and 9 different tunings, it quickly becomes clear how few 1,000 different sound snippets can be here.
The pricing is overambitious for what is offered
Soundraw charges about 17 EUR per month for an annual subscription paid in advance and about 20 EUR for a monthly subscription. That is – currently – a lot of money for what is offered.
By the way, I also noticed that you can’t delete your account on your own and there’s nothing in the FAQ about it. This is a so-called dark pattern and is supposed to make it difficult for uninvited users to get rid of this site again. After all, it is possible to easily cancel a paid subscription and remove payment means from the user account. After logging in, this is a maximum of 4 mouse clicks, which is absolutely fine. Nevertheless, Soundraw should make an improvement here and also enable account deletion by the user himself.
Let’s recapitulate my findings:
- A comparatively small library of source audio loops from 11 genres in 9 moods divided by different instrument genres at different speeds.
- Frequently repeating or very similar sounding results
- Comparatively high monthly costs
- Strongly limited licenses
- No easy way to delete user accounts
But also:
- Very easy to use
- Very quick generation of music tracks
- Very good sound quality
- Modern genres and modern instrumentation
- Very quick and competent support
From my side there is not yet a full recommendation
If you ask me, I doubt that there’s really an AI at work – there’s not enough evidence for that and I think the pricing is currently clearly overambitious for what’s on offer. Accordingly, I do not yet recommend a paid subscription. But you can checkout this by yourself. Soundraw offers a free 7 day membership.
If you are interested in deep learning and AI in music, you should take a look at “AIVA” instead – unlike Soundraw AI, AIVA is actually an intelligent system. Soundraw is cleverly made and can dazzle at first – but in fact it is just that: a dazzler!
Addition to the topic of support
Soundraw’s support earns top marks when contacted via email. Lightning fast, competent, honest and super friendly.
There’s currently a deal available:
69 EUR for a lifetime access to Soundraw. Depending on your trust into the company and the evolution of their service, this could be a good deal.
You’ll find more information here:
https://appsumo.com/products/soundraw/
This story will be updated soon as new findings show that Soundraw.io provides a very huge library with more than 550.000 single files and a size of roughly 195 GByte.
Currently, random checks are being made to see if there are duplicates among the loops, or if they are each rendered at different tempos and pitches, possibly also reused in different styles (which would significantly limit the variety despite the volume).
So please sty tuned and check this page frequently for getting updates
Hi Axel, do you have an update to this article? How are we feeling about Soundraw now? Thinking about subscribing but also noticed the repetitive remixes of similar tracks…
Hi,
according to my analyses at the time, there are a limited number of phrases that have been created with different instruments in some variations and the different pitches as finished sample snippets. Depending on the genre, there is sometimes more and sometimes less variation.
The total library at that time was about 200 Gbyte.
The arrangements were static. Depending on the genre, there was a maximum of about a handful of different arrangements, each tailored to a selectable length. If necessary, the arrangement was “extended” by repetition.
Rendering, according to my research, is done in the background using on Presonus Studio One. I was able to recover the corresponding arrangements as part of my analysis, so that post-processing in the music software was possible.
It should be noted in this respect that there is no evidence or even vague indications that AI is used in the generation of the music pieces. In this point I am almost 100% sure.
What is possible – but difficult to verify from the outside – is that the individual phrases are generated with the help of AI and post-processed by professional musicians before being rendered with different instruments in different pitches. An indication for this can be that the library grows only slowly. However, it is equally possible that the phrases are created or licensed in other ways.
I can’t make a decision for or against Soundraw for you. If you often need pleasing songs for presentation & co. and you don’t care how they are generated, Soundraw may be an option for you. If you really want to work creatively and are looking for inspiration, then AIVA might be worth a look 😉
I hope that I could help you.
Many greetings,
Axel
Thanks for this. I found this article while searching on soundraw because after my initial positive first impressions I began to feel very disillusioned with their service. I use them to generate music for daily content and after only a couple of weeks I began to notice that much of the music the site “generated” sounded very same-ish, with certain very distinct loops repeating in many selections. If, like me, you are staying within a certain set of genres and moods there is only a small amount of music to be had from this site before it becomes too repetitive to remain interesting and you should be prepared for your music too sound extremely similar to other users of the site.
Thank you so much for taking the time to share your thoughts and experience with Soundraw. It’s important to note that the number of available variations can vary greatly from genre to genre. I’ve conducted a comprehensive analysis in January 2023. The diagram attached below can help you understand the variation of different genres.
Thank you for your analyses! One thing I am wondering about is — it sounds like with Soundraw (as opposed to an actual AI), the user could feel more confident that they were not benefiting from AI’s scraping artists’ work from the web without their consent? Or does AIVA work differently?
It is important to understand that while Soundraw demonstrably does not use AI to generate the music tracks via the associated website at the time of my analyses – it is not possible to assess with certainty from the outside whether, for example, specialized AI is used to create the individual sound phrases (clips, patterns).
AIVA takes a completely different approach technologically. And this one seems (but I have not analyzed this in depth) to rely entirely on the use of ML/AI. There, complete pieces of music are generated on the basis of some parameters in the form of events (i.e. notes etc. – like in a MIDI file). It is even possible to upload your own music and create further compositions based on it – which also speaks for the use of ML/AI.
The problem you mentioned, that possibly copyrights of third parties are violated, unnoticed fragments of other compositions creep in, can in this respect (without deep analysis) neither with Soundraw nor with AIVA be completely excluded. Nevertheless, the risk of such rights violations is significantly lower with Soundraw due to the concept, since no complete compositions are created, but only individual phrases (clips, patterns) are combined into an overall arrangement. The rights infringement would therefore only be possible at the level of the clips/patterns – whether this would then still be legally relevant would have to be examined in each individual case.
However, I would like to point out that this is an abstract discussion. From the outside, it is not recognizable how and with what the models used have been trained. If they were trained exclusively with public domain or specially licensed data, there would be no legally relevant use.
Against this background, it is worthwhile to read the terms of use and licensing conditions of the providers carefully and to adhere to them. If the provider grants the corresponding rights of use, the buyer must be able to assume in good faith that everything is correct.
Great article and analysis… I your article after trying Soundraw for the first time and noticing the lack of variation and real controllability. Not only does this seem to not be “AI”, but its regurgitating identical tracks for different users! I decided to Shazam a couple of the tracks at random and BOTH OF THEM SHOWED UP AS EXISTING SONGS! Both songs were from 2023, so I’m guessing they originated as Soundraw tracks. One of them had an added lead vocal and the other probably doubled one of the instruments or added some other subtle element… But they sounded like the IDENTICAL (not just similar, but exactly the same) backing tracks to the “original” Soundraw versions “created” for me. I’m really curious as to what will happen when people start hearing “their tracks” being used by somebody else? Deceptive at best, but possibly illegal to claim “AI” created when it’s in fact not.
Thank you for your comment and the mention of Shazam. However, my analysis is solely focused on Soundraw itself and the way music is generated within Soundraw. After conducting an extensive reverse engineering project based on my initial analysis, I can provide accurate information about the scope of the phrase library, the supported genres, and their quantitative distribution in the pool. Furthermore, I have been able to completely replicate the generation of resulting musical pieces using my own program code.
Therefore, I can assure you that Soundraw doesn’t simply offer pre-made music tracks; rather, they are generated from a large construction kit of sorts.
As someone who produces my own music, I am aware that you can add your own music for later recognition on Shazam. It is possible that either Soundraw itself or individuals who have created their music using Soundraw have registered it with Shazam, leading to recognition. However, without further in-depth analysis, it is difficult to confirm or refute this possibility.
I was just trying to offer supporting info that it is definitely pulling from pre-written elements (as you showed)…
While they may be generated from a large “construction kit”, it is a small enough pool that when I Shazam’ed at random two pieces of music Soundraw “generated” for me, they showed up as complete identical songs which somebody else had already registered with Shazam. Please Shazam some tracks yourself and see if you get any results.
Point being, Soundraw represents that you’ll be getting some sort of unique piece of AI generated music…
If they had just said “hey, heres a royalty free library that you can customize online” they’d probably get some interest with that useful interface… but claiming it is “AI” is only a recipe for backlash. Its like a less cool sounding and less in-depth but easier to use (for the lay person) version of something like Splice’s online sequencer.
🙂
Thank you for the discussion. However, I believe you wouldn’t find as much joy with the Soundraw sound library. As mentioned earlier, it consists of 4-bar phrases available in different pitches, instrumentations, and variations. The extent varies depending on the genre, and you can find a breakdown in the comments.
At the time of my analysis, the preview library, accessible to non-paying customers, contained 1,114,000 files with a size of approximately 308 GB. The production library had 1,158,000 files and was around 1.58 TB in size. The differences stem from test material that varies across different pools.
Now, let’s return to Shazam. When I release a music piece, for example, through Ditto, Distrokid, or another aggregator, I can directly request song recognition in Shazam. Consequently, my music can be identified through Shazam. Shazam utilizes a kind of “fingerprint,” which simplistically represents a numerical mapping of a music piece. So, if someone were to release material from Soundraw, it would work in the same way.
Furthermore, I mentioned that based on my analysis, no AI is utilized in the web interface. However, this does not imply that AI is not used in any aspect. For instance, AI could be employed in the generation of the sound library, which is continually expanding. External verification or disproval is challenging, but it is plausible that Soundraw may employ AI at some point in the value chain.
Thank *you* for the discussion… I will just add one final response from my end 🙂
You are correct, I would not find much joy (if any) with the Soundraw library… My feeling on the quality of the music was that is not up to commercial standards, but interesting if AI created the full productions on the fly (as was implied in the ad thad led me to the product)…
Regarding Shazam… yes, that is correct what you are saying about Shazam… but my point there was that people will go to an AI music creation service because they want a totally original/new piece of music that was created specifically for them. And I used Shazam as a way to illustrate that Soundraw is giving multiple users not just similar music, but *identical* music based on what I found other people had released… and I was able to identify that music by using Shazam and finding identical pieces to what Soundraw had provided to me as “original”.
I’d be willing to wager that those people who’s songs I’d found through Shazam thought they were getting unique pieces of music from Soundraw created specifically for them through AI… i bet they’d be shocked to learn they just used a backing track from some “royalty free library”…I’m sure they thought they were releasing unique songs.
And regarding whether AI was used at any point in the process of creating these construction set loops… that is really immaterial. when you tell someone “choose your criteria and you will get an AI created song based on your choices!” you expect something created in real time for you…
Nobody uses AI software to get stuff somebody else made with AI and is giving the same product to multiple people. They use it to enter their own prompts and get a totally unique result.
From my side, I’d also like to add a final comment… I conducted a comprehensive analysis of Soundraw as part of a scientific case study for my studies. I even recreated the generation of music pieces using Python scripts. Unfortunately, the study is only available in German and it reveals significant flaws in Soundraw’s implementation. I’d like to give the creators adequate time to address these issues. The study includes detailed information about the arrangement format, quantitative aspects of the phrase library, variation diversity, and various implementation shortcomings. Perhaps you could check back occasionally – it’s possible that I’ll translate the study and publish it here at a later date.
But let’s get back to the core issue…
The problem with AI/ML applies not only to advanced systems like ChatGPT or DALL-E but also to Soundraw. As an end user, you can never be completely sure that third-party rights are not being violated. This is because you don’t know what the models were trained on, and there is no comparison of the generated content with corresponding archives.
In the worst-case scenario, plagiarism can occur. In the case of Soundraw, your observation can be easily explained. On one hand, the same original source could have been used by multiple people (or the same person) without permission, with someone paying for it and then sharing the results somewhere.
On the other hand, the music pieces generated by Soundraw are based on static, pre-made arrangements. Depending on the genre and desired length of the music piece, there are only a few variations of these arrangements (I can provide the exact numbers per genre if needed). Only the instrumentation and style variations are randomized using a random generator. This means that you can generate the same arrangement, chord progression, and instrumentation multiple times. It’s simply how Soundraw works. Mathematically, there could be millions of variations (although in practice, this is mostly theory, as sometimes only pitch and instrumentation vary, while the arrangement, chords, and melody remain the same ;-))
With my recreated generators, I can reproduce the same music piece as many times as needed or intentionally vary it, as I’m not limited to using a random generator.
If you only used the free version of Soundraw, it’s important to know that it uses compressed M4A audio files, which may not sound great (aside from the musical aspect). However, in the paid version, high-resolution WAV files are used, significantly improving the sound quality.
I’m not trying to convince you to become a customer of Soundraw. Nevertheless, I believe that the platform can be useful for certain purposes, such as prototyping games, generating music for inspiration, or providing background music for vacation videos or promotional films, where high musical standards may not be a priority. In these cases, whether AI is used or not becomes less important, and the usage license takes precedence. However, it is questionable when a provider advertises extensively with AI and yet there is no tangible evidence of it from the outside…