The words behind the signal.

Combining several imperfect sensors into one estimate that is better than any of them alone. Each sensor has a different failure mode, so a fusion filter weighs them against each other rather than trusting any single reading.

A recursive estimator that maintains a running best guess of a system's state along with how uncertain that guess is, then updates both as each new measurement arrives. It is the standard workhorse for fusing accelerometer, gyroscope and magnetometer data into a stable orientation.

Slow accumulating error in a quantity derived by adding sensor readings up over time. A small constant bias in an accelerometer becomes a growing velocity error, and then a much faster growing position error. Almost every hard problem in motion-based handwriting capture is a drift problem.

Adding acceleration up once to get velocity, then adding velocity up again to get position. It is the direct route from an accelerometer to a stroke path, and it is also why errors compound: any bias is multiplied twice over.

Resetting the velocity estimate to zero at a moment the system knows the tip is momentarily still, such as the start of a stroke or a pen lift. Each reset stops drift from carrying forward and keeps integration windows short.

Deciding moment to moment whether the pen is touching paper and inking or lifted and travelling. Much of a pen's motion is travel between letters and words, so getting this wrong draws phantom lines. It matters most for marks that sit away from the main stroke: the dot on an i, and the vowel signs and dots that hover above and below Indic aksharas.

A moment when the tip leaves the paper mid-word or between words. Counted and timed, pen lifts are one of the twenty handwriting factors, because frequent unplanned lifts indicate hesitation rather than flow.

How long the tip spends off the paper between strokes. It is invisible on a finished page and is one of the signals a motion-sensing pen adds that a scan of the same page cannot.

Recovering the path the tip actually traced from motion data alone. It has to overcome three things at once: drift from double integration, a gravity signal far larger than the writing signal, and the geometric fact that the sensors sit on the barrel rather than at the tip.

How writing speed rises and falls across a stroke. Skilled writers show a recognisable rhythm, quick along straight runs and slowing into tight curves, so the shape of the profile says more about control than average speed does.

The rate at which acceleration itself changes. Smooth, practised handwriting has low jerk; a hesitant or effortful hand produces sharp spikes, which makes it a useful measure of motor control.

The basic written unit of most Indic scripts. An akshara is not a single letter but a consonant, optionally combined with other consonants and a vowel sign, written and read as one block. This is why counting characters in Telugu or Kannada is not the same operation as counting letters in English.

Two or more consonants written as one combined shape rather than side by side. Conjuncts multiply the number of distinct forms a recogniser has to learn far beyond the size of the alphabet.

A subscript consonant mark written beneath the main consonant of a Telugu or Kannada akshara. Because it sits below the baseline and is often small, it is one of the parts generic OCR most reliably loses.

A vowel sign attached to a consonant in Telugu, changing the vowel the akshara carries. Many guninthams are written above or beside the base after the pen has lifted, which is exactly where accurate intent detection matters.

Turning an image of text into machine-readable characters. Mature for printed type, much harder for handwriting, and harder again for handwritten Indic scripts where labelled training data is scarce.

The handwriting-specific branch of OCR. Unlike printed type it has no fixed glyph shapes to match, so models learn the range of forms a character can take across many writers.

A neural network architecture built to read images by learning local visual patterns and combining them into larger ones. It is the standard approach for scoring letter formation and shape from a photograph of handwriting.

The gap between the data a model was trained on and the data it meets in use. Handwriting varies so much between writers, pens, paper and lighting that a model trained on one group of writers often fails on the next, which is a leading reason strong laboratory results do not survive real users.

Techniques for closing that gap without collecting a fresh labelled dataset for every new setting. Domain-adversarial training, for example, pushes the model to discard writer-specific quirks and keep only what distinguishes one character from another.

The verified correct answer a model is trained and measured against. For handwriting that means pages where a human has recorded exactly what was written, which is slow to produce and the reason Indic handwriting datasets stay small.