Technology
Everything SingDeep tells you about your singing traces back to a measurement, not a guess. If the evidence isn't there, it says so — it never papers over uncertainty with a global score or confident-sounding prose.
01 — The evidence chainFrom audio to advice
Every analysis flows through the same representation, and each layer keeps its provenance — which models ran, on what source quality, with what confidence:
A language model writes the coaching notes at the very end of this chain — but it only ranks and phrases what was already measured upstream. It can't invent a deviation that isn't in the data, or talk its way past one that is.
02 — The earWhat's under the hood
- Source separation — Demucs isolates the original vocal stem from any track, so you're compared against the singer, not the record.
- Pitch tracking — torchcrepe traces the f0 contour of both the reference vocal and your take, frame by frame.
- Alignment — dynamic time-warping over pitch, onset, and energy features lines up the two performances, with explicit ambiguity when paths compete.
- Expression features — onsets, vibrato rate/extent/delay, scoops and portamento, dynamics — the gestures that separate singing from note-reproduction.
- Coaching voice — an LLM translates the measured deviations into one concrete, specific next step.
03 — Measured, not vibesThe benchmark discipline
No component earns its place because a paper says it's state of the art. SingDeep maintains a frozen, deterministic benchmark corpus — synthetic contours with known truth: exact cents shifts, onset delays, time stretches, dropped phrases, scoops, vibrato. Every alignment or analysis change must beat the current baseline on that corpus before it ships.
Cases that are inherently ambiguous — expressive pitch, missing content, mixed changes — are labeled as such, and downstream reports are required to keep them uncertainty-qualified rather than collapsing them into a single correctness number.
04 — The flywheelHow it improves
Failures on the benchmark and disagreements with expert judgment become falsifiable experiments. A candidate method runs behind an adapter against the same immutable corpus, an independent review tries to refute the gain, and only surviving improvements integrate. Then the cycle repeats: measure, refute, integrate, verify.