Producers use AI stem separation to isolate vocals, drums, bass, and instruments from a mixed audio file, allowing them to sample, remix, or create karaoke tracks from finished songs.
Music production involves recording multiple individual audio tracks, such as microphones placed on a drum kit, a direct line from a bass guitar, and a dedicated microphone for a lead vocalist. When the recording process is complete, the mixing engineer balances all of these individual tracks and exports them as a single stereo audio file. For decades, once these tracks were mixed together into a master file, separating them again was practically impossible. Producers could use equalization to lower certain frequencies, but they could never truly isolate a single instrument from the mix. AI stem separation changes this process entirely by allowing producers to take a finished stereo mix and split it back into individual components. The technology relies on machine learning models that have been trained on thousands of hours of multitrack audio. By analyzing how different instruments behave in a complex mix, the software learns to identify and extract specific sounds. We use Demucs htdemucs, which is an open-source model developed from Meta AI research, to handle this process. We run this model entirely on our own servers to perform the audio separation efficiently.
What is AI stem separation and how does it work?
AI stem separation is the process of using artificial intelligence to divide a single audio file into multiple distinct tracks, which are known as stems. In traditional audio engineering, a stem is a sub-mix of multiple related tracks, such as all the drum microphones mixed down to one stereo pair. In the context of AI tools, a stem refers to the isolated output of a specific instrument group from a master recording. The process works by feeding the audio file into a neural network. The Demucs htdemucs model analyzes the waveform and separates the frequencies and phase information belonging to different instruments. Because we run this open-source model from Meta AI research directly on our own servers, the processing happens quickly. A 30-second preview usually comes in about a minute, and a four-minute song is fully split in about three minutes.
Where does AI stem separation come from?
Before artificial intelligence, audio engineers attempted to isolate sounds using phase cancellation. If an engineer had the exact instrumental version of a song, they could invert the phase of the instrumental and layer it over the original track to leave only the vocals. However, this required having the exact studio instrumental, which was rarely available to the public. Engineers also used heavy equalization to filter out low frequencies to remove a bassline, or high frequencies to dull the cymbals. This method always left artifacts and damaged the remaining audio. The shift occurred when researchers began training neural networks on large datasets of original multitrack sessions. The Meta AI research team developed the Demucs model to recognize the distinct sonic signatures of vocals, drums, bass, and other instruments, allowing for clean extraction without ever needing the original studio session files.
Why do producers use AI stem separation?
Music producers use these tools for several practical tasks. One common use is extracting vocals for a remix. By using an acapella extractor to pull the vocals out of any song, a producer can build an entirely new beat underneath the original singer. Beatmakers and hip-hop producers also use separation to sample older recordings. If a producer finds a sample but wants to remove the singer from it, they use an instrumental maker to make an instrumental from any song. This allows them to loop the backing track without vocal interference. Live performers and singers need backing tracks to practice or perform over. They often use a karaoke maker to turn any song into a karaoke track, or a vocal remover to remove vocals from a song with AI. Additionally, musicians learning to play an instrument use separation to hear exactly what the original player did. A bass player can isolate the bass stem to learn a complex riff, while a drummer can isolate the drums to study the groove.
How to perform AI stem separation with our tool?
To start the process, visit the Stem Splitter. You need a digital audio file of the song you want to split. You can upload an MP3, WAV, FLAC, M4A/AAC, or OGG song. The file must be up to 90 MB in size and up to 10 minutes long. The 90 MB file size limit accommodates high-quality uncompressed audio files. For example, a standard four-minute song in 16-bit, 44.1 kHz WAV format typically takes up around 40 to 50 MB of space, which fits comfortably within the 90 MB limit. The 10-minute duration limit ensures that the servers can process the files efficiently. This duration is long enough to handle almost all commercial music releases, including extended mixes and long progressive tracks, while preventing the system from being tied up by massive audio files. Links such as YouTube are not accepted, so you must have the actual file on your device. Users must have the right to process the audio they upload, and any copyright notices go to contact@stemsplitai.com. Splitting and previewing are free. You can process 1 song a day without an account. If you need to process more files, you can process 5 songs a day when signed in. Once you upload the file, the servers process the audio. Every split plays a 30-second preview of all five stems from 0:30 in the song. If you upload a song that is under a minute long, the preview will play from the start instead.
What stems and file formats do you get?
One split gives five stems. You receive the vocals, the instrumental, the drums, the bass, and the other stem. The instrumental stem is created by taking the drums, bass, and other stems and adding them together. The vocal stem contains all the lead vocals, backing vocals, and harmonies. The drums stem isolates the kick drum, snare drum, hi-hats, cymbals, and any other percussive elements in the recording. The bass stem extracts the low-frequency foundation of the track. The other stem collects the remaining melodic and harmonic content, such as guitars, pianos, synthesizers, string sections, brass, and any sound effects.
| Stem Name | Description |
|---|---|
| Vocals | The isolated singing or rapping |
| Instrumental | The drums, bass, and other added together |
| Drums | The isolated percussive elements |
| Bass | The isolated low-frequency elements |
| Other | Remaining instruments like guitars and keyboards |
When you download a whole song, you receive all five stems in two distinct audio formats. You get MP3 files at 320 kbps, which are compressed and take up less storage space. You also get WAV files at 16-bit, 44.1 kHz, which provide uncompressed audio for professional use. Your files remain private. Files are only available to the person who uploaded them. Uploads are deleted after 1 day, and the separated stems are deleted after 7 days.
| Format | Specification | Quality |
|---|---|---|
| MP3 | 320 kbps | Compressed |
| WAV | 16-bit, 44.1 kHz | Uncompressed |
How does the pricing and download system work?
While splitting and previewing are free, downloading the complete files requires credits. Downloading a whole song costs 125 credits, which is $0.50. This fee covers the download of all five stems in both MP3 and WAV formats. There is no subscription required to use the tool. You can make a one-time $5 purchase, which adds 1,250 credits to your account. This amount is enough for 10 songs. Credits are only spent on a download, so you do not lose credits when you upload a track just to hear the 30-second preview. All payments are processed securely through FloatPass. You can view more details on our pricing page.
Where else do people look for AI stem separation?
People searching for ways to separate audio tracks often explore multiple platforms and documentation to find the workflow that fits their needs. Besides our tool, people also look at sites like moises.ai and lalal.ai. For those using desktop software, manual.audacityteam.org provides documentation on audio editing. Other resources people check include help.moises.ai, isolate.audio, and podcast.adobe.com for various audio processing tasks.
Frequently asked questions
How much does AI stem separation cost on this site?
Splitting and previewing are free for 1 song a day without an account, or 5 songs a day when signed in. Downloading the full separated stems costs 125 credits ($0.50), and you can buy 1,250 credits for a one-time payment of $5.
What file formats work for AI stem separation?
You can upload MP3, WAV, FLAC, M4A/AAC, or OGG files up to 90 MB and 10 minutes long. Links from platforms like YouTube are not accepted.
How long does AI stem separation take?
A 30-second preview usually takes about a minute to generate. A full four-minute song is completely split into stems in about three minutes.
What stems do I get from a split?
One split gives you five stems: vocals, instrumental (which is drums, bass, and other added together), drums, bass, and other. You receive all five of these stems when you download the full song.
How long are my files kept on the server?
Uploads are deleted after 1 day and stems after 7 days. The files are only available to the person who uploaded them.
How do I use the preview feature for my tracks?
Every split plays a 30-second preview of all five stems from 0:30 in the track. If you upload a song that is under a minute long, the preview plays from the start.