percharacter

Percharacter, often written as per-character or character-level, is a term used in computing, linguistics, and typography to denote operations, metrics, or models that operate at the level of individual characters rather than words or subword units.

In text processing and OCR/ASR evaluation, per-character metrics include character error rate (CER) and character-level accuracy. CER measures mistakes per character, normalized by the reference character count. This approach is common for languages with rich morphology or non-Latin scripts, where word-level metrics may be less informative.

Character-level models process text as sequences of characters. They can capture spelling and long-range dependencies and are trained on raw text with minimal preprocessing. They contrast with word-level or subword models, and can generalize to unseen words or languages with limited word-level resources.

In typography and fonts, per-character information concerns glyph metrics, kerning, and rendering of individual characters. In user interfaces, per-character data underpins cursor movement, text selection, and accessibility features, ensuring precise interaction with text at the character level.

Limitations include longer sequence lengths and data sparsity for rare characters, as well as Unicode encoding challenges. Hybrid approaches that combine character-level signals with word- or subword-level representations are common to balance granularity with efficiency. See also character-level modeling, word error rate, and Unicode typography.