Audio embedding
A local CLAP-derived model represents how a file sounds and what language commonly describes it. This powers text-to-audio and audio-to-audio retrieval.
AI sample manager, without the cloud dependency
Your library already contains the answer. Ask it better questions.
Sample converts audio into a local 512-dimension semantic representation, combines it with BPM, key, duration, instrument, timbral metadata, and your text intent, then ranks the files already on your drives.
512floating-point dimensions in each local semantic audio vector
2 KBraw vector storage per analyzed file before index overhead
10 msdocumented text-embedding response target in the local helper architecture
0audio files intentionally uploaded for core search and analysis
“AI” should describe a workflow, not a sticker.
Sample uses different methods for different jobs. That makes behavior easier to explain—and easier to trust.
DSP analysis
Separate local analysis estimates BPM, key, duration, instrument family, and timbral properties. Confidence gates prevent weak metadata from pretending to be exact.
Intent parsing
Explicit phrases such as “120 BPM” or “C# minor” become structured constraints; descriptive language remains semantic. The two result channels are merged rather than conflated.
Vector search
A local approximate-nearest-neighbor index retrieves likely matches quickly. Filters and confidence-aware eligibility rules are applied before results reach the interface.
Inspectable, directional search
Use one sound as a starting point—then steer.
A reference sample anchors the search. “More” and “less” concepts change the vector direction. Optional BPM and key locks narrow the result set only when the reference and candidate metadata are reliable enough.
- Find similar
- More like
- Less like
- Lock BPM
- Lock key
What changes when retrieval becomes semantic?
The practical win is vocabulary independence: the pack vendor’s filename no longer has to match the phrase in your head.
Exact filename
narrow
Manual tags
medium
Audio similarity
broad
Text + audio + DSP
rich
Conceptual capability map, not a recall benchmark. Exact coverage depends on the query, library, metadata confidence, and model behavior.
AI boundaries matter.
A useful buying page should say what the system is not designed to do.
Retrieve by perceived character
Finds audio a person might describe similarly: material, mood, instrument, texture, density, brightness, impact, or production character.
Guarantee musical truth
Key and BPM are estimates with confidence rules. Similarity is not beat alignment, groove matching, copyright identification, or proof that two sounds are interchangeable.
Your original files
The catalog and analysis can change without renaming or moving audio. Normal library removal does not silently delete source files or preserved analysis.
Model the business case.
Use observed retrieval time from one real session for the most defensible estimate.
Adjustable workflow model
What is faster retrieval worth in your studio?
Your estimate
58 hr
recovered per year—about 1.2 hours per week, worth an estimated $2,042 of studio time.
Planning model, not a product-performance guarantee. It assumes 50 active weeks and uses only the values you enter. Time a normal session to replace the defaults with your own baseline.
One-time purchase · $129 launch price
Turn a folder archive into a creative search engine.
Sample searches and analyzes locally on Apple Silicon. Your audio stays on your Mac. Use it offline after activation, drag results into any DAW, and request a refund within 14 days if it does not improve your workflow.