From one microphone track, SonoDyno reads speed, cylinder count and load of an engine it has never heard

An engine recording without tachometer labels can now yield instantaneous speed and cylinder count directly, with microphone position, engine type and cylinder count all unknown. Reading speed from sound previously relied either on a known machine with a hand-chosen harmonic order or on general pitch trackers, which return one periodicity per frame and leave the count open. On two real engines absent from training, 93 to 97 percent of frames fall within 2 percent of the true speed, the cylinder count read once per recording session is 80 percent accurate, and fewer than 2 percent of frames suffer octave confusion.

The method turns the harmonic structure of engine sound into an energy map laid out by order, then lets the cylinder count enter the score only as weights. Order energy is computed on long, medium and short Blackman windows, each window's ridge is cleaned separately, and the windows are fused per order, so orders never mix before scoring. On the real validation recordings, even when each engine's harmonic order is supplied, the best general tracker, PESTO, places 61.6 percent of frames within a 5 percent tolerance against 95.7 percent for SonoDyno; with every learned stage removed and the three windows' order energy summed flat, that share falls to 72 percent.

Load is no longer replaced by loudness but estimated as an intensity relative to the engine's own capacity, 0 at the drag floor in overrun and 1 at the full-load ceiling. Engine sound synthesis trained on recordings previously used loudness as a proxy for load, and loudness carries neither the sign nor the scale of the torque. On a real engine of a known type that it has never heard, the load intensity has a mean absolute error of 0.12, and the sign of the torque, driving or braking, is right on 91 percent of frames.

The demonstrated scope is real engines recorded on a chassis dynamometer in a semi-anechoic room plus engines rendered by video-game engine simulations; real material covers only 4, 6 and 8 cylinders, the held-out real engines only 4 and 8, and every other count is tested on simulated engines alone. The intended reader is a researcher in sound synthesis, order analysis or noise and vibration who needs speed and load traces for unannotated engine recordings.

The remaining speed errors sit on ramps, in the fast-change stress band and at idle, where on real recordings roughly one frame in six to seven, one in three and one in six to seven is missed respectively. The one systematic miss reads a four-cylinder engine at constant high speed as eight, a confusion the signal model allows. On the unseen engine type the change of engine and the change of recording setup cannot be separated, and all real material was recorded in semi-anechoic rooms, so road, wind and traffic noise are not covered.