I hacked together a Python script to analyze the syllabic structure and pronunciation of Spanish text at the phoneme level, which it seems to get about 97% right, about one error every 30 words. It thinks each syllable consists of a (possibly empty) onset, a (nearly always empty) liquid, a vowel nucleus (which may be a diphthong), and a (usually empty) coda.
Here’s some example output, containing asterisks where it failed:
divide.^[173] /di-bi-de/ Los /los/ detalles /de-ta-ʃes/ del /del/ ciclo /si-klo/ celular /se-lu-laɾ/ solo /so-lo/ han /an/ sido /si-do/ investigados /in-bes-ti-ga-dos/ en /en/ el /el/ género /xe-ne-ɾo/ Sulfolobus, /sul-fo-lo-bus/ siendo /sien-do/ similares /si-mi-la-ɾes/ a /a/ los /los/ de /de/ bacterias /bak-te-ɾias/ y /i/ eucariontes: /e-u-ka-ɾion-tes/ los /los/ cromosomas /kɾo-mo-so-mas/ se /se/ replican /re-pli-kan/ desde /des-de/ múltiples /mul-ti-ples/ puntos /pun-tos/ de /de/ partida /paɾ-ti-da/ (origen /o-ɾi-xen/ de /de/ replicación) /re-pli-ka-sion/ usando /u-san-do/ ADN /ad-*/ polimerasas /po-li-me-ɾa-sas/ que /ke/ son /son/ similares /si-mi-la-ɾes/ a /a/ las /las/ enzimas /en-si-mas/ equivalentes /e-ki-ba-len-tes/ eucarióticas.^[174] /e-u-ka-ɾio-ti-kas/ Sin /sin/ embargo, /em-baɾ-go/ las /las/ proteínas /pɾo-tei-nas/ que /ke/ dirigen /di-ɾi-xen/ la /la/ división /di-bi-sion/ celular, /se-lu-laɾ/ como /ko-mo/ la /la/ proteína /pɾo-tei-na/ FtsZ /*-*-*-*/
It’s using mostly IPA, except for č (“ch”) and ñ. You can see it screwing up on a couple of words.
Based on this sort of analysis, it produces the following statistics, which I think are probably in the right ballpark, although it doesn’t recognize “ai” or “ui” as valid diphthongs, and the significant figures are mostly noise:
onsets: : 19.17% t: 11.42% d: 9.20% s: 8.86% k: 8.68% l: 7.93%
n: 6.30% m: 6.02% p: 5.56% b: 4.21% ɾ: 4.18% f: 2.35%
g: 2.03% x: 1.64% r: 1.14% č: 0.69% ʃ: 0.48% ñ: 0.14%
liquids: : 94.46% ɾ: 4.26% l: 1.28%
nuclei: e: 28.16% a: 26.47% o: 19.84% i: 12.79% u: 5.91% io: 2.16%
ia: 1.86% ie: 1.29% ue: 0.80% ua: 0.49% ei: 0.19% uo: 0.05%
codas: : 62.14% s: 15.65% n: 10.11% ɾ: 5.72% l: 2.40% m: 1.51%
k: 1.32% ks: 0.47% d: 0.20% p: 0.20% ns: 0.13% g: 0.08%
b: 0.05%
phones: e: 13.14% a: 12.44% s: 10.84% o: 9.52% i: 7.90% n: 7.14%
ɾ: 6.11% l: 5.01% t: 4.93% k: 4.52% d: 4.06% m: 3.25%
u: 3.13% p: 2.49% b: 1.84% f: 1.01% g: 0.91% x: 0.71%
r: 0.49% č: 0.30% ʃ: 0.21% ñ: 0.06%
This doesn’t capture all the structure of Spanish syllables; for example, only some onsets can be followed by liquids. But it is at least suggestive that the syllable structure is potentially exploitable for both encoding and human input methods (as, undoubtedly, stenographers already do.) The most common nucleus is /e/, followed closely by /a/, which between them capture the majority of syllables. The majority of syllables start with a vowel (19% empty onset) or either /t/, /d/, /s/, or /k/ (usually spelled “c” or “qu”), nearly all syllables lack a liquid following the onset, and the majority of codas are empty. So a syllable containing only “e” is 5.01 bits, one containing only “a” is 5.12 bits, the syllable “te” is 5.75 bits, and “tes” (as in eucariontes or equivalentes) is 7.74 bits. By contrast, a memoryless code represents “e” as 2.93 bits, more efficiently, but “tes” as 10.5 bits, because “t” almost never occurs at the end of a syllable, and “s” is much more common in the coda.
(Actually, my program doesn’t allow it to, so it can’t parse “habitat”, but it only very rarely encounters such a word.)
“bes” is 9.16 bits analyzed as a syllable, or 11.90 bits as three separate letters. “ti” is 6.87 bits as a syllable or 8.00 bits as two separate letters. “ga” is 8.31 bits as a syllable or 9.79 bits as two separate letters. So I guess often you could save about 20% of the effort that way.
However, the sort of long tail of diphthongs and multi-consonant codas is crushing my hopes for a human-written “relative positional” notation in which the same symbol is used for consonants in some places and vowels in others, and only the order of the symbols distinguishes between those meanings. I feel like it would just be too hard to learn given that all of the following occur in existing, if archaic, words: “crue” (“crueza”), “cruen” (“cruenta”), “cue” (“cuero”), “cuen” (“cuenta”), “que” (“que”), “quen” (“aquende”), “cu” (“culo”), “cun” (“secundaria”), and “cuns” (“circunscribir”, “circunstancia”). (I’m not totally sure bout “crueza” and “cruenta”, which I think might be three syllables.)
At the end of my program’s output it says:
bits per phone 3.879586262782223
bits per letter 3.586046123445401
This is 3.9 bits per phoneme (such as /ɾ/, /k/, or /e/) and 3.6 bits per letter (such as “h” or “c”), based on independently entropy-coding each letter. This is about 25% smaller than just flattening the usual Spanish alphabet, abcdefghijlmnñopqrstuvxyz, (to which we might reluctantly add k and w) because lg 27 ≈ 4.75 bits per letter, and the diaeresis for things like cigüeño and bilingüe; standardly we would also add áéíóúý, but I’m ignoring that here.
A pure Hamming phonetic code might assign 3-bit codes to /e/ and /a/, 4-bit codes to [soin], 5-bit codes to [ɾltkdmu], 6-bit codes to [pb], 7-bit codes to [fgx], an 8-bit code to [r] (“rr”), 9-bit codes to č [tʃ] “ch” and ʃ “ll”, and an 11-bit code to ñ. This would still leave 448 11-bit codes unassigned and would use 4.32 bits per phone or 3.99 bits per letter:
(- 2048 (+ 1 (* 4 2) (* 8 1) (* 16 3) (* 32 2) (* 64 7) (* 128 4)
(* 256 2))) ; 448
(+ (* 3 .1314) (* 3 .1244) (* 4 .1084) (* 4 .0952) (* 4 .0790) (* 4 .0714)
(* 5 .0611) (* 6 .0501) (* 5 .0493) (* 5 .0452) (* 5 .0406) (* 5 .0325)
(* 5 .0313) (* 6 .0249) (* 6 .0184) (* 7 .0101) (* 7 .0091) (* 7 .0071)
(* 8 .0049) (* 9 .0030) (* 9 .0021) (* 11 .0006)) ; 4.3196
(+ (* 1 .1314) (* 1 .1244) (* 1 .1084) (* 1 .0952) (* 1 .0790) (* 1 .0714)
(* 1 .0611) (* 1 .0501) (* 1 .0493) (* 1 .0452) (* 1 .0406) (* 1 .0325)
(* 1 .0313) (* 1 .0249) (* 1 .0184) (* 1 .0101) (* 1 .0091) (* 1 .0071)
(* 1 .0049) (* 1 .0030) (* 1 .0021) (* 1 .0006)) ; 1.0001
However, whether for human interface stuff or for data compression, there’s no reason to restrict yourself to radix 2 nowadays; IBM’s arithmetic-encoding patents are long expired, and I’m typing this on a 100-key keyboard, though the home row and the keys above and below are only about radix 33. Still, it’s a good ballpark that [ea] should be the easiest to write, followed by [soinɾltkdmu], and then [pb], requiring about twice as much effort as [ea], then [fgxr], then “ch” (č) requiring about three times as much effort as [ea], then finally [ʃñ].
You might omit č entirely, just spelling it out as /tʃ/, although that would be pretty rough on speakers of other dialects of Spanish, even worse than lumping [z] in with [s].
If we wanted to divide this up in some sort of rational and easy-to-learn way, we might try to express [pbfgxrčʃñ] as some kind of modified version of [soinɾltkdmu], which could be divided up into vowels [oiu], sonorants [lmnd], plosives [ɾtk], and the sibilant [s]. (Spanish /d/ has a fricative allophone [ð] and a plosive allophone [d].) There’s inevitably going to be significant tension between fluent writing and ease of learning.
Here's the Python program:
#!/usr/bin/python3
"""Analyze phonetics and phonotactics of Spanish text.
The syllabification is somewhere around 97% accurate on the text I
tried it on, but it does have problems, for example with “mayoría”
(should be ma-yo-ri-a; same problem with “podrían”, “amoníaco”, and
“biotecnología”) and “inusuales” (should be in-u-sua-les). Also it
thinks “hay” and “muy” are two syllables each. It fails on
“habitats”, “atmosférico”, “plancton”, and “sulfhidrico”. And it
segments “aire” as “a-i-re”, which I think is wrong, and similarly
“causado” as “ca-u-sa-do”, "autodenominado" as “a-u-to-...”,
“contribuir” as “con-tri-bu-ir”, and “cuidadores” as “cu-i-da-do-res”.
No attempt is made to handle numbers, initialisms, or loanwords not
spelled using Spanish orthography.
It’s using a slightly deformed version of IPA. For [tʃ] I’m using the
Esperanto “č”, in the interest of having it be one letter for the
purpose of frequency tabulation, and for [ɲ] I’m using the Spanish
“ñ”, in the interest of readability. No attempt is made to select the
appropriate allophone for context (e.g., [ŋ] for /n/, [β] for /b/, [x]
for /s/). “*” is output on error. The pronunciation is mostly
Argentine. Here’s some sample output, containing one error:
> divide.^[173] /di-bi-de/ Los /los/ detalles /de-ta-ʃes/ del /del/
ciclo /si-klo/ celular /se-lu-laɾ/ solo /so-lo/ han /an/ sido
/si-do/ investigados /in-bes-ti-ga-dos/ en /en/ el /el/ género
/xe-ne-ɾo/ Sulfolobus, /sul-fo-lo-bus/ siendo /sien-do/ similares
/si-mi-la-ɾes/ a /a/ los /los/ de /de/ bacterias /bak-te-ɾias/ y /i/
eucariontes: /e-u-ka-ɾion-tes/ los /los/ cromosomas /kɾo-mo-so-mas/
se /se/ replican /re-pli-kan/ desde /des-de/ múltiples /mul-ti-ples/
puntos /pun-tos/ de /de/ partida /paɾ-ti-da/ (origen /o-ɾi-xen/ de
/de/ replicación) /re-pli-ka-sion/ usando /u-san-do/ ADN /ad-*/
polimerasas /po-li-me-ɾa-sas/ que /ke/ son /son/ similares
/si-mi-la-ɾes/ a /a/ las /las/ enzimas /en-si-mas/ equivalentes
/e-ki-ba-len-tes/ eucarióticas.^[174] /e-u-ka-ɾio-ti-kas/ Sin /sin/
embargo, /em-baɾ-go/ las /las/ proteínas /pɾo-tei-nas/ que /ke/
dirigen /di-ɾi-xen/ la /la/ división /di-bi-sion/ celular,
/se-lu-laɾ/ como /ko-mo/ la /la/ proteína /pɾo-tei-na/ FtsZ
/*-*-*-*/
Here’s some more, containing three errors in 110 words:
> El /el/ hecho /e-čo/ tuvo /tu-bo/ tintes /tin-tes/ mafiosos
/ma-fio-sos/ y /i/ elementos /e-le-men-tos/ que /ke/ ya /ʃa/ se /se/
habían /a-bian/ visto. /bis-to/ Lo /lo/ protagonizó
/pɾo-ta-go-ni-so/ un /un/ grupo /gɾu-po/ de /de/ encapuchados
/en-ka-pu-ča-dos/ autodenominados /a-u-to-de-no-mi-na-dos/
“mapuches” /ma-pu-čes/ en /en/ Río /rio/ Negro, /ne-gɾo/ el /el/
domingo /do-min-go/ a /a/ la /la/ noche. /no-če/ Primero /pɾi-me-ɾo/
ataron /a-ta-ɾon/ a /a/ los /los/ cuidadores /ku-i-da-do-ɾes/ de
/de/ un /un/ predio /pɾe-dio/ de /de/ Vialidad /bia-li-dad/
provincial, /pɾo-bin-sial/ después /des-pues/ dejaron /de-xa-ɾon/
notas /no-tas/ intimidatorias /in-ti-mi-da-to-ɾias/ con /kon/
amenazas /a-me-na-sas/ y, /i/ antes /an-tes/ de /de/ escapar,
/es-ka-paɾ/ incendiaron /in-sen-dia-ɾon/ un /un/
depósito. /de-po-si-to/
> El /el/ Gobierno /go-bieɾ-no/ rionegrino /rio-ne-gɾi-no/ calificó
/ka-li-fi-ko/ como /ko-mo/ “un /un/ acto /ak-to/ terrorista”
/te-ro-ɾis-ta/ el /el/ ataque /a-ta-ke/ y /i/ prepara /pɾe-pa-ɾa/
una /u-na/ presentación /pɾe-sen-ta-sion/ ante /an-te/ la /la/
Justicia /xus-ti-sia/ federal. /fe-de-ɾal/ En /en/ paralelo
/pa-ɾa-le-lo/ pidió /pi-dio/ al /al/ Gobierno /go-bieɾ-no/ nacional
/na-sio-nal/ el /el/ envío /en-bio/ de /de/ fuerzas /fueɾ-sas/
federales /fe-de-ɾa-les/ para /pa-ɾa/ controlar /kon-tɾo-laɾ/ la
/la/ situación /si-tua-sion/ que /ke/ viene /bie-ne/ escalando
/es-ka-lan-do/ desde /des-de/ hace /a-se/ meses. /me-ses/ Se /se/
suma /su-ma/ el /el/ incendio /in-sen-dio/ provocado /pɾo-bo-ka-do/
en /en/ la /la/ Oficina /o-fi-si-na/ de /de/ Turismo /tu-ɾis-mo/ de
/de/ El /el/ Bolsón, /bol-son/ ocurrido /o-ku-ri-do/ el /el/ sábado
/sa-ba-do/ por /poɾ/ la /la/ noche. /no-če/
<https://www.clarin.com/sociedad/mapuches-incendiaron-campamento-vialidad-rio-negro-gobernadora-pidio-apoyo-gobierno-nacional_0_ByRHc8tKZ.html>
Despite the problems mentioned earlier, I think it’s accurate enough
for these statistics to be mostly right:
onsets: : 19.17% t: 11.42% d: 9.20% s: 8.86% k: 8.68% l: 7.93%
n: 6.30% m: 6.02% p: 5.56% b: 4.21% ɾ: 4.18% f: 2.35%
g: 2.03% x: 1.64% r: 1.14% č: 0.69% ʃ: 0.48% ñ: 0.14%
liquids: : 94.46% ɾ: 4.26% l: 1.28%
nuclei: e: 28.16% a: 26.47% o: 19.84% i: 12.79% u: 5.91% io: 2.16%
ia: 1.86% ie: 1.29% ue: 0.80% ua: 0.49% ei: 0.19% uo: 0.05%
codas: : 62.14% s: 15.65% n: 10.11% ɾ: 5.72% l: 2.40% m: 1.51%
k: 1.32% ks: 0.47% d: 0.20% p: 0.20% ns: 0.13% g: 0.08%
b: 0.05%
phones: e: 13.14% a: 12.44% s: 10.84% o: 9.52% i: 7.90% n: 7.14%
ɾ: 6.11% l: 5.01% t: 4.93% k: 4.52% d: 4.06% m: 3.25%
u: 3.13% p: 2.49% b: 1.84% f: 1.01% g: 0.91% x: 0.71%
r: 0.49% č: 0.30% ʃ: 0.21% ñ: 0.06%
bits per phone 3.879586262782223
bits per letter 3.586046123445401
"""
from __future__ import print_function, division
import collections, math, re, sys
syllable = re.compile(r'''
(?P<syllable>
(?P<onset>b|c|ch|d|f|[gq]u(?![lr])|g|h|j|k|ll|l
|m|n(?![lr])|ñ|ph|p|rr|s|th|t|v|y|z|)
(?P<liquid>l|r|)
(?P<nucleus>a|ei|e|w|ia|io|ie|i|o|ua|ue|u|y)
# The negative lookahead assertions are a hack to keep
# from chomping up codas that really belong to the
# onset of a following syllable: na ve gac ion,
# en cic lo ped ia, etc.
(?P<coda>(?:b|c(?!h)|d|g|l|m|ns|r|p|s|z)(?![aeiouylr])|x|n(?![aeiouy])|)
)
| (?P<err> .)
''', re.VERBOSE)
accents = {'á': 'a', 'é': 'e', 'í': 'i', 'ó': 'o', 'ú': 'u',
'ü': 'u'}
def syllabize(word):
w = ''.join(d for d in (accents.get(c, c)
for c in word.lower())
if d == 'ñ' or 'a' <= d <= 'z')
return list(syllable.finditer(w))
def pronounce(syllable, first):
if not syllable.group('syllable'):
yield '*'
return
onset = syllable.group('onset')
liquid = syllable.group('liquid')
nucleus = syllable.group('nucleus')
coda = syllable.group('coda')
mapped = {'ch': 'č', 'c': 'k', 'qu': 'ku', 'gu': 'gu', 'h': '',
'll': 'ʃ', 'y': 'ʃ', 'ph': 'p', 'th': 't', 'v': 'b',
'z': 's', 'rr': 'r', 'r': 'ɾ', 'j': 'x',
}.get(onset, onset)
if nucleus[0] in 'ei' and not liquid:
yield {'c': 's', 'g': 'x', 'gu': 'g', 'qu': 'k'}.get(onset, mapped)
else:
# Treat “guas” from “aguas” as the same “ua” nucleus as “tuan”
# as “interactuando”:
if mapped.endswith('u'):
mapped = mapped[:-1]
nucleus = 'u' + nucleus
# Treat “rra” from “tierra” the same as “rra” from “rápido”:
if liquid and not onset:
mapped = 'r' if liquid == 'r' and first else 'ɾ' if liquid == 'r' else liquid
liquid = ''
yield mapped
if liquid == 'r':
yield 'ɾ'
else:
yield liquid
yield {'w': 'u', 'y': 'i'}.get(nucleus, nucleus)
yield {'c': 'k', 'z': 's', 'x': 'ks', 'r': 'ɾ', 'j': 'x'}.get(coda, coda)
def pronounce_word(word):
return [list(pronounce(s, i == 0))
for i, s in enumerate(syllabize(word))]
def print_counter(name, counter):
print("%8s:" % name, end='')
total = sum(v for k, v in counter.items())
nl = False
for i, (k, v) in enumerate(counter.most_common()):
if nl:
print(' ' * 9, end='')
nl = i % 6 == 5
print('%2s: %5.2f%%' % (k, 100 * v / total),
end='\n' if nl else ' ')
if not nl:
print()
def main(stdin):
onsets = collections.Counter()
liquids = collections.Counter()
nuclei = collections.Counter()
codas = collections.Counter()
phones = collections.Counter()
letters = 0
for line in stdin:
for word in line.split():
p = pronounce_word(word)
print(word, '/%s/' % '-'.join(''.join(s) for s in p), end=' ')
if not any(s == ['*'] for s in p):
for s in p:
onsets[s[0]] += 1
liquids[s[1]] += 1
nuclei[s[2]] += 1
codas[s[3]] += 1
for c in ''.join(s):
phones[c] += 1
letters += len(word)
print()
print()
print_counter('onsets', onsets)
print_counter('liquids', liquids)
print_counter('nuclei', nuclei)
print_counter('codas', codas)
print_counter('phones', phones)
t = sum(v for v in phones.values())
entropy = sum(-phones[c]*math.log(phones[c]/t)/math.log(2) for c in phones)
print("bits per phone", entropy/t)
print("bits per letter", entropy/letters)
if __name__ == '__main__':
main(sys.stdin)