Click Run Analysis and STEMWORKS turns a folder of tracks into sections, tempo maps, keys, tags, and exportable segments.
STEMWORKS runs the recording through a fixed set of stages. Each stage produces one part of the final timeline.
A folder of WAV files from your multitrack recorder or DAW. STEMWORKS handles chunked recordings and stereo pairs. It uses the master or mix track when one exists, or builds a mono mix from all tracks.
A timeline of segments with time ranges, tempo estimates, keys, instrument tags, and song-likeness scores. You see the same data in the multitrack timeline and in the Songs table for scanning and batch export.
Everything runs locally. Your audio files stay on your machine. STEMWORKS does not upload stems or require an internet connection to analyze a session.
The pipeline has six stages. Segmentation runs first because the rest of the work depends on those boundaries. The remaining stages run in parallel.
Splits the full recording into meaningful time ranges — songs, transitions, interludes, and silence. This is the foundation for everything else. See how it works below.
Estimates the BPM, musical key, and chord progression for each segment. Long segments are analyzed in smaller windows to handle tempo changes. Beat positions are saved for snap-to-beat editing later.
Measures loudness inside each segment to decide which tracks are playing. The adaptive threshold keeps one loud hit from knocking quieter tracks out of nearby segments.
Listens to each track’s audio in each segment and tags what instrument is playing — drums, bass, guitar, keys, vocals, etc. When the full recognition model isn’t available, a simpler method is used and labeled as lower confidence.
Builds multi-resolution peak data for every track. You can zoom from the whole session down to individual beats without reloading raw audio.
Writes the results to a local database with provenance, run times, and source-audio fingerprints. That lets STEMWORKS resume or re-run analysis without starting from zero.
Incremental and resumable. Each stage caches results from the input fingerprint. Re-run the same files and STEMWORKS skips unchanged work. If the app stops mid-run, it resumes at the first incomplete stage.
Segmentation decides where one section ends and the next begins. STEMWORKS ships three methods, with structural segmentation as the default.
This method compares every moment of the recording against the rest of the jam to find repeated material and section changes. Matching chord cycles or grooves cluster together. A new riff or a section break shows up as a boundary.
You get cuts based on musical structure: repeated parts, section starts, and clear shifts in the jam.
The audio is converted into a representation of its harmonic content over time. Every point is compared against every other point to build a similarity map. Repeating structures (like a verse appearing twice) show up as patterns in this map.
The similarity map is scanned for transition points where the music changes character. You can raise or lower sensitivity depending on how many boundaries you want to catch.
Detected boundaries are organized at three levels: fine-grained phrases (~2 seconds), musical sections (~5 seconds), and song-level groups (~30 seconds). This multi-scale view means STEMWORKS can find both small details and large structures.
Each segment gets a score from 0 to 1 for how song-like it is. Segments with strong repeated structure score high. Long stretches of noodling score low. STEMWORKS uses that score for song and transition classification.
This method cuts on silence and abrupt level changes. It runs in under 3 minutes regardless of jam length. Use it for a rough first pass or for sessions where structural segmentation takes too long.
Runs structural and energy-based segmentation, then merges both boundary sets. This catches silence gaps, musical transitions, and other cases where one method sees a split that the other misses. Very short fragments are collapsed afterward.
Tune without re-running. After analysis completes, adjust the song detection thresholds in the filter bar. STEMWORKS reclassifies segments on the spot. You do not need another analysis pass.
Each segment gets its own tempo, key, and chord estimate. If the jam speeds up or changes key, those sections keep their own values.
Tempo is analyzed directly from the primary audio source (the master track or mix). For segments longer than a minute, the analysis runs in shorter windows and combines the results to handle drift. Beat positions are saved so you can snap segment boundaries to the nearest beat when editing.
Each segment gets a set of key candidates ranked by confidence, along with a detected chord progression. The analysis looks at the harmonic content — which notes and chords are being played — and estimates the most likely key and progression. Silent segments are flagged as unknown rather than producing misleading results.
Once the segments are established, STEMWORKS figures out which tracks are active in each one and what instruments they contain.
Each track’s loudness is measured in each segment. The adaptive threshold compares each track against its own neighbors, so quiet instruments still register when they are present. A hard floor excludes silent tracks.
In the timeline, you can see which tracks played in each segment. STEMWORKS also uses that data when it builds segment exports and participation counts.
STEMWORKS listens to each track’s audio within each segment and tags the most likely instrument — drums, bass, guitar, keys, vocals, and more. These tags appear in lane tooltips, segment details, and are included in export metadata. When the full recognition engine isn’t available, a simpler method runs instead and the result is labeled as lower confidence so you know the difference.
Once analysis finishes, you can move the results into a DAW or hand them to collaborators.
If the Ableton integration is running, STEMWORKS can create locators and import clips directly into a Live set with correct tempo and warp settings. Without the integration, it exports a segment pack (WAV files + metadata) with step-by-step import instructions.
Exports a transfer package with a MIDI timeline (tempo, time signature, markers) and full-length WAV stems. The Import Wizard guides you to drag the MIDI file in first to set the tempo map, then place the WAV stems on the grid.
Generates .dawproject packs for Studio One and Bitwig, including the tempo, arrangement, and audio stems needed to rebuild the segment in your DAW.
Export an Advanced Authoring Format (AAF) file when you need multitrack audio placed on an absolute timeline for post or video work.
Mark segments as loops or one-shots, then batch-generate WAV artifacts. Grid-aware loop export uses detected transients and downbeats so the files land closer to the grid.
Export any segment as an MP3 with fades and embedded metadata for sharing rough cuts or keeping an archive.
The timeline shows per-track waveforms, segment overlays, BPM and loudness graphs, and transport controls. The Songs view gives you a sortable table for rating, naming, tagging, and batch export.
Desktop playback reads the original track files through the native audio engine. Browser playback uses preview proxies generated in the background. The master track plays at once, and individual proxies show a “Warming” badge until they are ready.
The hybrid analysis algorithm processes approximately 15 GB of audio per hour on a modern laptop. Structural segmentation takes 10–30 minutes for 1–3 hours of audio; energy-based segmentation completes in 1–3 minutes. Processing time varies with CPU speed, disk speed, and recording complexity. The progress panel shows per-stage status and an estimated time remaining.
If you re-run without changing your files, STEMWORKS skips unchanged stages. Adjusting song detection thresholds does not trigger another analysis pass.
You can switch jams, close the app, or let your machine sleep during analysis. STEMWORKS resumes from the last unfinished stage on the next run. Clear analysis resets the run.
WAV, AIF, and AIFF audio files from DAW exports or standalone hardware recorders (e.g., Zoom LiveTrak, H-series, Tascam machines). Multi-file chunks, stereo pairs, and various sample rates are handled automatically. macOS 12+ and Windows 10/11.
The desktop app starts with a 14-day trial. Import, analyze, explore, and test limited exports with your own recordings. No credit card required.