Stenotype keyboard layout generator
Using some optimization algorithms, evolve a stenograph keymap layout that minimize complexity of the theory for a given language.
Stenograph keyboards present a limited keyset where multiple keys are pressed at the same time, forming a chord that represent one or more phoneme / syllable in a word. Steno machines were once proprietary and expensive, but cheaper options are now available through custom keyboards using popular mechanical keyboards parts and software or firmware interpreter. Those new keyboards are programmable, so it is not needed to stick to the original century-old phoneme-keymap layout.
Visit the Open Steno Project for an in depth introduction to the subject.
For hobbists, there is little added benefit to learning a traditional steno layout if better layouts can be generated. Steno keyboards are not common and it is unlikely that one will need to use a steno keyboard that is not his own.
This project aims at generating keymaps and theories for any keyboard layout (number and arrangement of key switches) based on some keyboard constrains (prefered keys) and using a .
Stenotyping is done on custom (minimal keys) keyboards. The user simultaneously presses multiple keys to form syllables or groups of syllables. The fingers move as little as possible, but each can press multiple keys situated on the same column, or adjacent columns to the pinky or index fingers.
Those keypress (or chord) are interpreted by a software layer implementing a Theory. The theory translate the chord or group of chords into one or multiple words.
The theory must closely match the syllable(s) pressed to the typed word(s), otherwise learning and remebering the corresponding keypress will be difficult. It must also distinguish between words that have the same pronunciation, but differnent spellings (homophones).
The best theory for a lexicon is as easy to learn and use as possible.
The algorithm will generate multiple keymap and theory pairs for a given keyboard layout, physiognomical constrains and lexicon. Those keymap-theory will be scored based on the strain and complexity in the following ways :
The average number of keystroke must be minimized :
- Chords containing less keystrokes are preferable
- Words containing less Chords are preferable
- Common words should contain less strokes than rarely used words
- Movement of fingers should be minimized
- Pressing fewer keys per finger is preferable
The mapping of keys to the sound they represent must be coherent :
- Phonemes must have a maximum of one canonical representation on the keyboard, one portion of a chord (todo: validate)
- Syllables must have the minimum number of chord representation (variations) on the keyboard to distinguish between the different spellings
- Rules (chord variations) to distinguish between spelling of a syllable must be consistent across a maximum of words sharing that syllable
- As much as possible, the order in wich phonems are typed must match the order they occure in the syllable
Words that do not respect an established rule are deemed an exception
- The number of exceptions must be minimized
The French language Lexique version 383 was used as a base [1] [2]. This lexicon version provides Word orthograph, phonology, Consonant-Vowel breakdow and phonological syllable breakdown (ex.: Manger - To eat, m@Ze, CV-CV, m@-Ze). Most importantly, it offers the words frequency extracted from 2 differnt corpus : Written French books and French movies subtitle. It is however incomplete in regards to the grapheme (spelling) breakdown of syllable with only 2/3 of the words providing a breakdown allowing to map phonetic syllables to written syllables.
A second lexicon, LexiqueInfra [3], was used to build graphem representations of the syllable. This lexicon provides phoneme-grapheme
associations for 137k of the 142k words of Lexique383. The final lexicon resources/LexiqueMixte.tsv was built by lexique.py and contains
136,348 French words with corpus frequencies and both phoneme and grapheme syllable breakdowns.
Typically, the order of the phonemes (from left to right) typed to form a chord must ressemble the order of the phonemes in the syllable. Ordered biphoneme frequencies informs on the order the phonemes should be placed on the keymap. For this example, we assumed that the typical order of the keys on the keymap would be the 3 groups of phonemes : Left-consonants -> Vowels -> Right-consonants. Syllables in French do not typically have two groups of vowels phonemes separated by consonants (VCV), so this mapping is consistent with the typicals French syllables (V, VC, CV, CVC).
While biphoneme frequencies inform on the order of the phonemes, single-phoneme frequencies inform on the importance of the phoneme. Here is a representation of a phoneme order for all 3 groups of phonemes that minimize the frequency of chords where the phones are in the wrong order. Bar charts represent the frequency of the individual phonemes in each of the 3 groups.
Left hand optimization :
Best order (ordered score 91215.3, disordered score -1400.5):
dZksNptgfvSmnbzRljw
> R | <
> R | <
> R | <
> t R | <
> s pt R | <
> s pt m R | <
> d ks pt v m Rl | <
> d ks pt v m Rlj | <
> d ks pt fv mnb Rlj | <
> dZks ptgfvSmnbzRljw | <
> dZksNptgfvSmnbzRljw | G <
^^^^^^^^^^^^^^^^^^^^^^|^^^^
ordered | floating
Right hand optimization :
Best order (ordered score 15522.8, disordered score -5759.7):
wpjfvbskdgtRNlzmnZS
> R | <
> R | <
> R | <
> R | <
> R | <
> R | <
> R | <
> s tR | <
> s tR l | <
> skd tR l | <
> wpjfvbskdgtRNlzmnZS | G <
^^^^^^^^^^^^^^^^^^^^^^|^^^^
ordered | floating
Vowel optimization :
Best order (ordered score 6585.4, disordered score 0.0):
8§O92aoE@5ie
> a | <
> a | <
> a e | <
> a E ie | <
> a E ie | <
> a E ie | <
> a E@ ie | <
> aoE@ ie | <
> § aoE@ ie | °uy <
> §O aoE@5ie | °uy <
> 8§O92aoE@5ie | °uy1 <
^^^^^^^^^^^^^^^|^^^^^^^
ordered | floating
Negative (disordered) score represent the sum of frequencies of biphonemes that would be in the wrong order. In the case of the right hand consonants, the 25% disorded to 75% orderded ratio is mainly due to to the "R" biphonemes where "R" can be before or after other consonant phonemes:
- "tR" frequency: 4110
- "Rt" frequency: 1822
- "dR" frequency: 2265
- "Rd" frequency: 1339
The algorithm choose the least penalizing option, placint "R" after "t" and "d".
Work is ongloing in keyboardtemplate.py where a description of the Starboard keyboard is provided. The different
keypresses are assigned to fingers and penalty scores loosely corresponding to the strain they induce.
This is the last step to obtain an optimal phonetic keyboard. Most popular keys must be easily accesible. Phonemes that require key combos must not cause conflicts with other words that use the same keys to represent other phonemes.
Typicaly, homophones have "alternative spelling" where extra phonemes or the * key are added to distinguish the different ortographical spellings. A system must be devised to make the treatment of these exceptions so that it is intuitive
French verbs have multiple conjugations suffixes per tense, some of which are homophones (je mange - m@Ze, tu manges - m@Ze). A system based on the function of the verb may simplify the selection of verb homophones.
Other common prefixes and suffixes should have consistent phoneme-keymap associations to reduce the congitive load to lear exceptions.
[1] New, B., Pallier, C., Brysbaert, M., Ferrand, L. (2004) Lexique 2 : A New French Lexical Database. Behavior Research Methods, Instruments, & Computers, 36 (3), 516-524.
[2] New, B., Brysbaert, M., Veronis, J., & Pallier, C. (2007). The use of film subtitles to estimate word frequencies. Applied Psycholinguistics, 28(4), 661-677.
[3] Gimenes, M., Perret, C., & New, B. (2020). Lexique-Infra: grapheme-phoneme, phoneme-grapheme regularity, consistency, and other sublexical statistics for 137,717 polysyllabic French words. Behavior Research Methods. doi.org/10.3758/s13428-020-01396-2