Meitei Mayek handwriting recognition stops slicing words into characters: whole-word reading cuts synthetic test word errors to 1.57%
Synopsis
Handwritten Meitei Mayek has been recognised one character at a time, yet its most frequent confusions depend on where a character sits in a word. Training words are composed from isolated character images and read whole by a convolutional encoder, a bidirectional LSTM and CTC, giving 0.24% character and 1.57% word error on 5,000 synthetic test words; on 100 unseen real words written on a tablet by one native writer, character error is 2.96% and word error 16.0%.
Interpretation
Handwritten Meitei Mayek can now be read as whole words rather than segmented into characters and classified one at a time. Prior work on this script concentrated on isolated characters, where an ensemble of six convolutional networks reaches 98.12% on the official TUMMHCD test set, and no published work recognises handwritten Meitei Mayek words or lines end to end. On 5,000 synthetic test words the whole-word recogniser with a character language model reaches 0.24% character error and 1.57% word error (mean of two seeds); the same characters cut from the words with perfect boxes and given to that ensemble with perfect zones and the same language model give 4.26% character error and 20.98% word error.
Training words are composed from isolated character images instead of transcribed handwritten words. TUMMHCD's 85,124 character images are all 24 by 24 pixels with ink scaled to fill the frame, so size, aspect ratio and position within a word are lost, and the database carries no writer information. Positions come from the Noto Sans Meetei Mayek font, sizes from stroke thickness, and spacing from six published handwriting samples where 33.5% of neighbouring letters touch and the rest sit 0.097 L apart; in the synthetic words 32.4% of neighbouring letters touch, with visible gaps of 0.059 L.
The recogniser reads a whole word image as a left-to-right sequence without character boundary labels. A ConvNeXt-T encoder initialised from a TUMMHCD character network feeds a bidirectional LSTM and CTC, and decoding adds a character 6-gram language model. The encoder keeps the first three ConvNeXt-T stages and halves only the height at the third, giving 4 rows with one column per 8 pixels; on the 100 real handwritten words the final recipe with the language model reaches 2.96% character error and 16.0% word error, and 118 of its 119 character errors involve the three signs that print places above the letter.
Perspective
The demonstrated scope is evaluation on synthetic words plus 100 real words written in digital ink on a tablet by one native writer and text-disjoint from training; the code, models, real word set and a browser demonstration are publicly available.
The real set covers one writer, one device and 100 words, and the placement of the signs above was measured on this one hand, so whether other writers place them the same way is not known; the synthetic results are in-distribution and on their own say little about real handwriting; TUMMHCD has no writer identities, so synthetic words mix writers within a word and no synthetic split can be writer-disjoint.