When it comes to artificial intelligence in music production, there is seemingly a new product dropping every week, each promising to offer something groundbreaking. From LALAL.AI’s stem separation and reverb remover tools to ethically trained MIDI generation plugins, AI in music production has progressed in leaps and bounds in recent years.
For some people, AI-powered music production tools are considered to be controversial due to questionable AI model-training practices. These tend to be the tools that are trained on others’ music without credit or consent, and are designed to help the layperson create fully-fledged musical compositions.
Other types of AI music production tools are less contentious. Some are intended to simply streamline a process, or in the case of Audialab’s Deep Sampler 2 (DS2), generate the basic building blocks with which producers can create their own compositions. Today, we’re reviewing DS2 and taking a deep dive into its function, its operation and ultimately, whether it works.
What is Deep Sampler 2?
Audialab’s DS2 is the second iteration of their text-based AI-powered sample generation plugin. If you’re wondering if that means what you think it means, the answer is yes: You feed the plugin a text prompt, it does some thinking, and then generates an audio sample that you can drag straight into your DAW.
The basic version of DS2 is currently in Alpha and is completely free. There’s also a paid Pro version, which adds batch generation functionality along with the ability to create and share plugin-like effects.
Audialab’s Approach to AI
Deep Sampler isn’t Audialab’s only foray into artificial intelligence. Their Humanize plugin uses AI to add natural variation to your drums, and Emergent Drums 2 works similarly to Deep Sampler, but is specifically designed to generate drum sounds.
It’s clear that Audialab’s goal isn’t to rob talented musicians of a career, but rather to equip them with innovative new tools that allow them to harness their creative talent in exciting new ways. There’s even a dedicated page on their website that outlines their stance on ethical AI in music production.
How Does Deep Sampler 2 Work?
Let’s examine DS2’s operation in a bit more detail.
Deep Sampler 2’s Three Components
Currently, there are three main components to DS2. One is the plugin itself, which comes in AU and VST3 formats. The second component is the AI model used to generate the audio from your text prompt, and the final component is the Audialab engine, which runs as a standalone application.
Downloading AI Models
Before you start generating audio, you’ll need to download at least one of the three currently available models. While DS2 does have an integrated model downloader, it wasn’t working during our testing, with downloads freezing at 50% or not starting at all. Instead, we had to manually download the models from their respective Hugging Face repositories.
Currently, there are only three readily available models that work with DS2. These are Infinite Pianos, Vocal Textures and Stable Audio Open, the latter being a general-purpose model for generating a range of sounds.
Configuring Deep Sampler 2’s Settings
With your models downloaded, an instance of the DS2 plugin loaded in your DAW and a connection established between it and the Audialab Engine, you’re almost ready to generate some audio.
You first need to select one of your downloaded models from the Model List drop-down menu at the top left of the plugin interface. You can also configure DS2’s engine to refine the audio being generated, but we’ll kick things off with the default settings.
Generating Audio
With DS2 fully configured, you can now type a positive prompt in the prompt bar. This essentially dictates to the engine the kind of sound you would like it to generate, for example ‘deep distorted kick drum’. You can also add keywords in the negative prompt bar to tell the engine what sounds you don’t want it to generate. This is useful for steering the model away from incorrectly generating sounds that resemble something else.
Before hitting the Generate button, you can also set the tempo and length of the sample. Longer samples may contain more variation, but will also take longer to generate. The first time you choose a model and generate a sample, DS2 will need to load the selected model. This can take a few minutes, but once loaded, the sample generation itself takes less time. The time it takes depends on how you’ve configured DS2, but generally doesn’t seem to take longer than 10 to 15 seconds.
Once DS2 has finished generating your sample, you can click the audio waveform at the top of the plugin interface to preview it. You can then go back and tweak the engine’s settings to refine the sound. In our experience, though, using the settings to generate the exact sounds in your imagination isn’t always possible.
Is Deep Sampler 2 Any Good?
The Good
DS2 is capable of generating some truly compelling sounds, particularly with the dedicated Piano and Vocal models which are trained specifically on those instruments. Once you’ve loaded a model, it can create relatively realistic sounds in only a matter of seconds. The sounds generated by DS2 don’t sound inherently digital, unlike some other AI-powered audio-generating tools. They sound deep and rich, with plenty of frequency and stereo content.
Even when the plugin doesn’t quite get it right, and instead generates a completely different sound, the result is often still quite musical and very useable. For example, when trying to generate a ‘tight breakbeat in a big space’, DS2 gave us an evolving FM synth-like tonal pad. While we didn’t receive what we ordered, we did get a sound that would have taken some time to design manually, and could easily work in a range of productions.
And why is that useful? DS2 helpfully stores all of your previously generated sounds under the Library tab, neatly categorised by the model used to create them. So even if you don’t quite get the sound you’re after, you’ll start to build up your own bespoke library of unique samples for future use. What’s more, you can instantly preview each sample in your library with the Play button.
The Bad
It’s important to note that Audialab are really pushing the boat out with DS2, which is still in Alpha. This means there are of course some quirks and limitations to the application of this technology. Between the DS2 plugin, the Audialab engine and any models you want to use, Deep Sampler 2 consumes a considerable amount of storage. With all three models installed on our device, the whole package is around 20GB.
Again, what Audialab have done with DS2 is incredibly novel, so things aren’t always as smooth as you might like them to be. As mentioned, we struggled to download the models in DS2’s integrated downloader. Even once we’d downloaded the models, loading them took a few minutes each. While this might be down to our specific device, DAW or plugin build, it did slow things down for us. That said, these were the only teething issues we had, and DS2 behaved itself overall.
Finally, the current option of only three models feels slightly restrictive. As mentioned, the Vocal and Piano models yield considerably stronger results than the general-purpose model. This left us yearning for more advanced models for other instruments, such as strings, bass and guitar.
It also seems odd that there’s no dedicated drum model, but perhaps this was a conscious choice from Audialab given their dedicated drum sample generation plugin. On the other hand, the basic version of DS2 is completely free, and the possibility of adding further models as they become available is very appealing. On the DS2 product page, Audialab say there are “dozens more [models] coming soon”.
The Verdict
Audialab DS2 is a seriously promising piece of music software for turning text prompts into bespoke samples. The dedicated instrument models are particularly impressive, and the musicality of their output means DS2 should be able to easily find a place in your production process. Even if you’re not looking for vocal or piano samples, DS2’s ability to conjure up intriguing new sounds that you wouldn’t have otherwise considered makes it useful for sparking inspiration too.
The few teething issues we had with downloading and loading models are to be expected from a tool as ambitious as this. While it was slightly frustrating to get up and running, it was smooth sailing once the initial issues were ironed out.
Ultimately, using DS2 is a fun and engaging experience, which is exactly what music production should be. While DS2 rarely nails a prompt at the first time of asking, it’s more satisfying when it does eventually get it right. Adjusting DS2’s settings and prompts to hone in on the sound in your imagination feels more like a fun problem-solving exercise than a struggle, and can throw up some interesting sounds and textures that would be difficult to create manually.
If you’ve got enough hard drive space, we highly recommend you give Audialab DS2 a try. It is completely free, after all.
🍿🍿 Read our other reviews:
