Trace the entire path
The signal passes through the sensor, its polling cycle, control-data processing, the audio buffer, and the converter. Reducing only the buffer size is not enough if the sensor updates infrequently or the gesture map is heavily smoothed.
Processing speed and response time are different characteristics. An algorithm may process a second of audio in less than a second, yet wait for a long input segment before starting. That may be acceptable for file processing; in live performance, the musician will feel the delay after a strike. Measure SOMATIC’s control-data path and REGOLITH’s audio path separately, then check their combined response.
Distinguish average latency from variability
A small, constant delay feels different from a randomly changing one. For a rhythmic interface, it is useful to measure event timing spread as well as the average delay.
An average hides occasional long responses. Record the number of trials, median, an upper percentile, and the largest observed value. The maximum in a short test is not a guaranteed worst case. Compare an idle setup with the usual load: interface open, recording active, and every required channel running. That is the operating condition the musician will encounter.
Measure the complete action
A practical test compares the physical instant of impact with the output audio signal. Record the measurement conditions: sample rate, buffer size, load, and sensor settings. Without them, figures are difficult to compare.
Define gesture onset and sound onset before measuring. A contact impulse can timestamp a strike, whereas the start of a gradual breath depends on a chosen threshold. Document that threshold and the detector. An electrical loopback can measure the audio interface but excludes sensor waiting time and mechanical travel. These tests answer different questions and should be recorded separately.
Choose the musical trade-off
A long spectral-analysis window helps distinguish nearby frequencies, but adds latency. That can be acceptable for a cloudlike decay effect; a direct strike may need a separate fast path.
At 48 kHz, a block of 64 samples spans approximately 1.33 ms. That is one block’s duration, not a promised hand-to-loudspeaker latency: queues, converters, and the algorithm add their own contributions. GRAIN can retain a fast direct attack alongside a slowly developing layer. For SEREIN, an intentional delay can become part of the phrase if its behavior remains understandable to the performer.
What the research adds
Using neural synthesis, Caspe and colleagues show why real-time operation does not guarantee low latency. Their analysis relates response to buffering and model architecture. We use it to frame measurements, not as a specification for REGOLITH or SOMATIC. [1] Caspe et al., 2025
Takeuchi and Saito study jitter in audio sampling instants. Its scale and mechanism differ from variation in musical-event response time. Sharing the word “jitter” does not make the measurements interchangeable. [2] Takeuchi & Saito, 2023
they often suffer from high latency
A practical check
Keep a short set of consistently performed strikes and slow changes, then repeat it after changing the buffer size. Record sample rate, software versions, sensor mode, and load. Compare the numbers and articulation under the same monitoring conditions. For TESSERA’s mechanical strikers, account separately for travel time: prompt command delivery does not mean an immediate acoustic impact.
Research references
Links point to the specified arXiv versions; publication status is available on each paper’s page. The studies concern their authors’ systems. Connections to Sagittarius A Studio instruments are our engineering and musical interpretation.