Part of our guide to transcribing a full band.
When a recording has one instrument doing one thing, transcribing it is straightforward. The trouble starts with a full mix, where vocals, guitars, bass, and drums all occupy the same moment and the same frequencies. Separating the recording into stems first, one file per part, gives you a clean version of each instrument to transcribe on its own. This is not always necessary, but when parts are masking each other it makes the difference between a transcription that tracks the music and one that smears it together. Here is what stem separation does, when it is worth the extra step, the tools that do it, and how to skip it when you can.
What stem separation is
Stem separation splits a finished mix back into its parts. A model trained to recognize each kind of source pulls the vocals onto one track, the drums onto another, the bass onto a third, and groups the rest, leaving you with a small set of audio files instead of one blended recording. It is not flawless; you can usually hear a little bleed from neighboring instruments in each stem. But a part on its own is far easier to hear and to transcribe than the same part competing inside the full mix. If the idea is new to you, what stem separation is covers how the models work in more detail.
When it helps, and when to skip it
Reach for stem separation when the mix is dense and the part you want is hard to pick out by ear, because if you cannot hear it cleanly, a transcription will struggle too. A wall of guitars over a busy rhythm section is the classic case. Skip it when the recording is already sparse, like a solo piano take or a single vocal line, where splitting adds a step without adding clarity. Songscription does not require stems at all: you can transcribe a full mix and concentrate on one instrument, so treat separation as a tool you bring out for crowded recordings rather than a mandatory first step. The trade-off between doing the whole band at once and one part at a time is covered in per-instrument vs one-pass transcription.
How to separate stems
Several tools do this well, and they fall into two camps. Open-source models such as Demucs and Spleeter run on your own machine and are free, which suits a one-time setup and batches of songs. Web tools like LALAL.ai and Moises split a track from the browser with nothing to install, which suits a quick one-off. Whichever you use, the workflow is the same: feed in the full song, get back separate files for vocals, drums, bass, and the rest, and listen to the one you plan to transcribe to confirm it is clean enough to work from. Judge a tool by the part you actually care about, not the average across all the stems.
Transcribing the stems into a score
Once you have clean stems, transcribe them one at a time. Run the vocal stem through Songscription to get the melody, the bass stem to get the bass line, and so on, so each part is worked out against a clear signal instead of a crowded one. Then combine the parts into a single score in a notation program. This per-part approach is the whole strategy behind the multi-instrument transcription guide, and if you want the details of any single instrument, the instrument transcription guide covers piano, guitar, bass, horns, voice, and drums in turn.
Transcribe a clean part at a time
Separate the mix if you need to, then upload each part and get an editable score back. The free tier is enough to transcribe your first stem.
Frequently Asked Questions
Do I need to separate stems before transcribing a song?
Not always. For a solo instrument or a sparse recording, transcribe the audio directly. Stem separation helps when a mix is dense and several instruments overlap, because isolating one part at a time gives the transcription a cleaner signal to work from. Songscription does not require stems: you can transcribe a full mix and focus on one instrument, and reach for stem separation only when parts are masking each other.
What does stem separation do?
Stem separation takes a finished mix and splits it back into its parts, typically vocals, drums, bass, and everything else, using a model trained to pull each source out of the blend. The result is several audio files, one per part. It is not perfect; you often hear faint bleed from other instruments, but a separated part is far easier to transcribe than the same part buried in the full mix.
Which tool should I use to split stems?
Several do it well. Open-source models like Demucs and Spleeter run locally, and web tools such as LALAL.ai and Moises split a track in the browser. Any of them produces the isolated parts you then transcribe. Pick by whether you prefer something local and free or a quick upload, and judge the result by how clean the part you care about sounds.
How do I get a full score after separating stems?
Transcribe each stem on its own, then combine the parts. Run the vocal stem, the bass stem, and so on through Songscription one at a time to get a clean part for each, then assemble them into a single score in a notation program. Transcribing per part rather than all at once is what keeps each line accurate, which is the approach the multi-instrument guide is built around.
The fastest way to start is on the part that matters most in your song. Upload a recording with Songscription and transcribe one clean part at a time.
