Speed up Smart Fades beat tracking with a numba-accelerated decoder

[marcelveldt]

What does this implement/fix?

The Smart Fades beat/downbeat tracker ran its DBN (HMM/Viterbi) post-processing as a pure-Python per-frame loop, which pegged a CPU core for the full duration of every track's analysis and contributed to sluggishness while streaming. This adds a compiled (numba) decoder that produces bit-identical beats/downbeats while running ~4-5x faster, reducing the per-track CPU spike.

• Add a numba-jitted Viterbi decoder (nogil, compiled once at provider init) used by default, with the existing numpy decoder kept as an automatic fallback and readable reference implementation.
• Precompute the per-meter decode plan (state classification, predecessor tables) once at init instead of rebuilding it for every track.
• Declare numba in the Smart Fades manifest (already present transitively via the core librosa dependency).
• Add tests asserting the numba and numpy decode paths are bit-identical to each other and to a naive reference Viterbi.

Related issue (if applicable):

• N/A

Types of changes

• Bugfix (non-breaking change which fixes an issue) — bugfix
• New feature (non-breaking change which adds functionality) — new-feature
• Enhancement to an existing feature — enhancement
• New music/player/metadata/plugin provider — new-provider
• Breaking change (fix or feature that would cause existing functionality to not work as expected) — breaking-change
• Refactor (no behaviour change) — refactor
• Documentation only — documentation
• Maintenance / chore — maintenance
• CI / workflow change — ci
• Dependencies bump — dependencies

Checklist

• The code change is tested and works locally.
• pre-commit run --all-files passes.
• pytest passes, and tests have been added/updated under tests/ where applicable.
• For changes to shared models, the companion PR in music-assistant/models is linked.
• For changes affecting the UI, the companion PR in music-assistant/frontend is linked.
• I have read and complied with the project's AI Policy for any AI-assisted contributions.
• I have raised a PR against the documentation repository targeting the main or beta branch as appropriate.

[github-actions]

🔒 Dependency Security Report

📦 Modified Dependencies

music_assistant/providers/smart_fades/manifest.json

Added:

• ✅ numba >=0.51

Unchanged dependencies

• beat-this ==1.1.0
• nnAudio ==0.3.3

The following dependencies were added or modified:

diff --git a/requirements_all.txt b/requirements_all.txt
index 17ac2851..be7f45fd 100644
--- a/requirements_all.txt
+++ b/requirements_all.txt
@@ -53,6 +53,7 @@ music-assistant-models==1.1.129
 mutagen==1.47.0
 niconico.py-ma==2.1.0.post2
 nnAudio==0.3.3
+numba>=0.51
 numpy==2.3.5
 orjson==3.11.6
 pillow==12.2.0

New/modified packages to review:

• numba>=0.51

---

🔍 Vulnerability Scan Results

No known vulnerabilities found

Name	Skip Reason
torch	Dependency not found on PyPI and could not be audited: torch (2.11.0+cpu)
torchaudio	Dependency not found on PyPI and could not be audited: torchaudio (2.11.0+cpu)
✅ No known vulnerabilities found

---

Automated Security Checks

• ✅ Vulnerability Scan: Passed - No known vulnerabilities
• ✅ Trusted Sources: All packages have verified source repositories
• ✅ Typosquatting Check: No suspicious package names detected
• ✅ License Compatibility: All licenses are OSI-approved and compatible
• ✅ Supply Chain Risk: Passed - packages appear mature and maintained

Manual Review

Maintainer approval required:

• I have reviewed the changes above and approve these dependency updates

To approve: Comment /approve-dependencies or manually add the dependencies-reviewed label.

[Copilot]

Pull request overview

This PR accelerates Smart Fades’ DBN/HMM (Viterbi) beat/downbeat post-processing by adding an optional numba-jitted decoder and by precomputing decode-plans per meter at tracker initialization, aiming to reduce per-track CPU spikes during analysis.

Changes:

• Add a numba-accelerated Viterbi decoder with a numpy fallback path.
• Precompute per-meter decode plan (state classification / predecessor tables) once at init instead of per-track.
• Add tests to assert bit-identical outputs across numba vs numpy vs a naive reference implementation, and add numba to dependencies.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
music_assistant/providers/smart_fades/dbn_postprocessor.py	Adds numba-jitted decoder, decode-plan precompute, and switches runtime decode to a numba-first path with numpy fallback.
music_assistant/providers/smart_fades/manifest.json	Declares numba as a Smart Fades provider requirement.
requirements_all.txt	Adds numba to the aggregated dependency list used in CI installs.
tests/providers/smart_fades/test_dbn_postprocessor.py	Adds correctness tests ensuring numba/numpy/naive Viterbi decoders are consistent and output-identical.

[MarvinSchenkel]

Just be aware that the gains are not substantial. Postprocessing takes around 5% of the the total inference time, so we are optimizing on a small portion already. For example, total inference time on my prod setup takes like 30s with 2s being the post processing time.

[marcelveldt]

Let this one for you to decide @MarvinSchenkel - in general a C-backed library is always better than pure python so maybe good to do some comparison - no rush.