Neuropsychologia 45 (2007) 425–430

Note

A dedicated neural mechanism for vowel selection: A case of
relative vowel deficit sparing the number lexicon
Carlo Semenza a,c , Giulia M.L. Bencini b,∗ , Laura Bertella a,c,e , Ileana Mori c,d , Riccardo Pignatti a,c ,
Francesca Ceriani c , Danielle Cherrick b , Emanuela Magno Caldognetto f
a University of Trieste, Italy
b Department of Psychology, Hunter College, City University of New York, 695 Park Avenue, New York, NY 10021, United States
c Istituto Auxologico Italiano, Verbania, Italy
d University of Bergamo, Italy
e Clinica Hildebrand, Brissago, TI, Switzerland
f Istituto di Scienze e Tecnologie della Cognizione, CNR, Fonetica e Dialettologia, Padova, Italy

Received 29 November 2005; received in revised form 27 June 2006; accepted 8 July 2006
Available online 25 September 2006

Abstract
We report the case of an Italian speaker (GBC), with classical Wernicke’s aphasia following a vascular lesion in the posterior middle temporal
region. GBC exhibited a selective deficit in spoken language production affecting vowels more than consonants. In reading from a newspaper, GBC
substituted vowels for other vowels from the Italian inventory at a rate of 7/1 compared to consonants. No effects of frequency or grammatical
class were found. Vowel errors could also not be accounted for by morphological or known phonological processes. Production of number words,
in contrast, was free from phonological errors. While GBC has intact representations of Italian vowels and consonants, his data argue for a separate
selection mechanism for vowels that is dissociable from that used for consonants. This case provides neuropsychological evidence for models of
word production that distinguish between the abstract phonological properties of a word (e.g., sequencing of phonemic slots, or “CV skeleton”)
and a separate representation for the specific sounds (melody).
© 2006 Elsevier Ltd. All rights reserved.
Keywords: Vowels; Consonants; Aphasia; Vowel production; Number words

A pervasive finding in the literature on phonological errors
in speech is that on average, aphasics and normal speakers make
more errors on consonants than on vowels (Béland, Caplan, &
Nespoulous, 1990; Blumstein, 2001; Canter, Trost, & Burns,
1985; see Dressler, Tonelli, & Magno Caldognetto, 1990, for
Italian). The incidence of selective deficits on vowels in aphasia
is unknown, but, like many selective deficits, we can expect it
to be low, and therefore, virtually undetectable in group studies
where individual differences between the patients are often lost.
The general aphasiological literature (reviews can be found
in Blumstein, 2001; Denes, Semenza, & Magno Caldognetto,
1999; Nickels, 2001) has focused on broad phonological differences between classic aphasic syndromes, and relatively little
attention has been devoted to the phonemic level. Studies that

∗ Corresponding author.

E-mail address: gbencini@hunter.cuny.edu (G.M.L. Bencini).
0028-3932/$ – see front matter © 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.neuropsychologia.2006.07.006

have dealt with the neural basis of phoneme production generally
assume a neural network distributed in perisylvian areas.
Based on patterns of phonological errors in aphasics, a model
proposed by Blumstein (2001) distinguishes between a lexical
selection and a phonological planning stage before articulatory
implementation. On this model, patients with conduction aphasia would have difficulty in the planning stage, while those with
Wernicke’s aphasia would struggle at the lexical selection stage.
Conduction aphasics frequently approximate the sound structure of the target word: their phonemic substitutions seldom
differ from the target in more than one distinctive feature. It
is believed that they correctly select the lexical target but fail in
properly realizing the phonetic details of the word. In contrast,
the production of Wernicke’s aphasics often does not closely
resemble the phonological structure of the target, suggesting
that the problem lies in selecting the correct target word among
competing lexical candidates. However, as is the case with most
phonological studies in aphasia, this model is based on data from

426

C. Semenza et al. / Neuropsychologia 45 (2007) 425–430

consonant substitutions and does not deal in any detail with the
production of vowels.
There are only two published studies on patients exhibiting selectively impaired performance on vowels compared to
consonants: case MM (Romani, Granà, & Semenza, 1996),
and case AS (Caramazza, Chialant, Capasso, & Miceli, 2000).
These studies were conducted in repetition. Both patients
were native Italian speakers with left-temporo-parietal vascular
lesions resulting in typical conduction aphasia. The case studies with MM and AS do not address whether their deficit is
one of vowel representation or vowel selection. They also fail
to examine the respective contribution of articulatory, prosodic,
and lexical factors on error patterns.
The present investigation reports a third case of relative vowel
deficit, with some important additional findings. It provides a
more in-depth investigation of the nature of the errors, revealing
an intact abstract representation for the phonological skeleton
of a word (CV skeleton), and a deficit in selecting the specific
vowel from an equally intact language specific vowel inventory. Additionally, this patient produced number words without
phonemic deficits. Dissociations resulting in intact phonological
performance on number words in the face of impaired phonology
on other word classes have been reported before (e.g., Cohen,
Verstichel, & Dehaene, 1997). We will argue that the number
lexicon may be accessed by a separate non-semantic route in
which number words are retrieved as sequences of preassembled, phonetically spelled-out, units.
1. Case report
At the beginning of this investigation, GBC was a 62-yearold, still active and successful business man, with only 5 years of
formal education. He had a long history of vascular problems. In
1996, he suffered and immediately recovered from a transitory
ischemic episode with language problems affecting the production side. In February 1998, he suffered an ischemic lesion in
the left posterior temporal area, resulting in fluent aphasia of
the classical Wernicke type. An old ischemic lesion in the left

occipital lobe was also found on that occasion. In May 1998, he
underwent a bypass between the superficial temporal artery and
the middle cerebral artery. The procedure was complicated a few
days after the intervention by bleeding in the temporo-parietal
area.
In December 2004, GBC was evaluated at the Neurology
Department of the Istituto Auxologico Italiano of Verbania,
where the present investigation took place from January to May
2005. An MRI taken in April 2005 revealed a vascular lesion in
the posterior middle temporal region (Fig. 1). Aphasia was the
only detectable neurological impairment.
GBC’s full neuropsychological evaluation also revealed significant symptoms for aphasia only. His language functions were
evaluated via a battery of tests including the Italian version of
the Aachener Aphasia Test (Luzzatti, Willmes, & De Blaser,
1996), and portions of the Miceli, Laudanna, and Burani (1994)
battery for the assessment of aphasia in Italian. GBC’s Wernicke’s aphasia was stable during the investigation and was distinguished by well-articulated, relatively fluent, paraphasic and
paragrammatic speech, severely affected repetition and naming, and moderately impaired auditory comprehension. In each
production task, it became obvious that the majority of single
phoneme errors affected vowels rather than consonants. This
unusual observation led to the present investigation.
GBC’s native language was Italian, in the Piemontese variant spoken by communities in the west side of Lago Maggiore
(Province of Verbania). He competently spoke standard Northern Italian with the experimenters, with a regional Piemontese
accent. GBC collaborated enthusiastically throughout the investigation.
2. Experimental investigation
2.1. Collection of spoken material
GBC was tape-recorded in all experimental tasks. Three judges, all native
speakers of GBC’s dialect independently transcribed the tapes. GBC’s articulation was very clear. No more than a couple of words were coded as unintelligible,
and these were due to voice lowering. Inter-judge agreement was close to 100%.

Fig. 1. MRI showing posterior middle temporal lesion in case GBC.

C. Semenza et al. / Neuropsychologia 45 (2007) 425–430

2.2. Background testing
This section reports in detail a number of tests to determine the best task to
examine the vowel relative to consonant deficit. Unless otherwise stated, these
tasks were taken from the Miceli et al. battery (1994). All tasks used an even
number of high and low frequency targets where applicable.
2.2.1. Picture naming
In picture naming of objects and actions (half low and half high frequency items), GBC made 40/58 errors, again mostly consisting of non-targetrecognizable neologisms; the few phoneme substitutions concerned vowels only.
2.2.2. Repetition
In repetition of single words (15 nouns, 15 verbs, and 15 function words,
all three/four syllables long), GBC made 32/45 errors, mostly neologisms with
little to no relation to the target: only five single phoneme substitutions were
found, four of which concerned vowels.
2.2.3. Phonemic discrimination and spoken-to-written syllable matching
Phonemic discrimination and spoken-to-written syllable matching tasks
were administered to determine whether GBC’s repetition errors were due to
a phonological comprehension problem. Both tasks contrasted pairs of CV syllables, presented one at a time. (Consonants were occlusives, e.g., PA, TA, CA,
BA, DA, GA; answer yes/no). GBC showed a mild deficit; he made 7/60 errors
on the phonemic discrimination task, and 6/60 errors when matching spokento-written syllables.
2.2.4. Lexical decision (“is this a real word or not?”)
In the auditory modality, GBC made 13/40 errors on real words (10 nouns,
10 adjectives, 10 verbs, and 10 function words), and 1/40 errors on non-words.
In the visual modality, he only made 2/40 errors on real words and 4/40 on
non-words.

427

words. This shows that, for GBC, if a number word was contained in a larger
unit (whether this was a real word or not) it was not processed as a number word,
and therefore it was not immune from vowel errors.

2.3. Interim discussion
Little is known about complex number reading, but one possibility is that
large numbers are parsed into smaller sequences of more frequent basic number
words (e.g., four, thousand, five, hundred, etc.) (Cohen et al., 1997).
Basic number words in Italian (number words for digits 1–19, tens, hundreds, thousands, million(s) and the suffix – mila, – thousand) are one to four
syllables in length and vary in stress pattern (e.g., “trè”—three, “ùno”—one,
“ùndici”—eleven). They contain 94 vowels and 110 consonants (vowel rate
85%). The distribution of vowels within the basic Italian number lexicon is
comparable to the distribution of vowels within the general Italian lexicon. The
frequencies of vowels /a/, /e/, /i/, /o/, /u/, in the basic number lexicon are as
follows: (cf. frequency in the entire lexicon in parenthesis): /a/ 20 (25), /e/ 20
(28), /i/ 34 (20), /o/ 16 (20), /u/ 10 (7).
Number words have a restricted number of possible intra-category competitors and, indeed, GBC’s only errors were “lexical” substitutions; he made
virtually no number “syntax” violations. His lexical errors consisted of substituting the whole lexical unit with another lexical unit, and could not be accounted
for by contextual vowel substitutions.
Even if complex numbers are read by parsing them into smaller more frequent units, GBC’s performance cannot be explained on a frequency factor: as
mentioned above, and shown also in reading connected speech, he made vowel
errors even on extremely high frequency function words, such as prepositions
and articles.
Therefore, the preservation of number words in this patient does not appear
to stem from particular lexical or phonological properties of number words (such
as frequency or distribution of phonemes), rather it stems from the possibility
of accessing number words via non-semantic routines (see Section 3).

2.4. Reading connected prose
2.2.5. Lexical discrimination
To assess the receptive aspect of reading, GBC was given 40 pairs of written
words, 4–10 letters long: 20 pairs were the same word written in a different font
(“casa” – house – “CASA”), the other 20 pairs consisted of words differing by
one letter only; either a consonant or a vowel (“RAMO” – branch – “remo”
– oak). He was flawless in deciding whether the two words were the same or
different.
2.2.6. Reading aloud
In single word reading (52 nouns, 20 verbs, 20 function words—words were
half high and half low frequency), GBC made 10/92 errors, 8 of which were
vowel substitutions. He was less efficient in connected prose, where the rate of
phonological errors was similar to that found in spontaneous speech (where he
produced some neologisms but more semantic paraphasias and anomic hesitations). No between-word errors were observed.
2.2.7. Word class and length effects: reading and repetition of numbers
While there was an overall word length effect, no clear frequency, concreteness, or grammatical category effects appeared. The category of numbers
however, was an exception. GBC not only had a digit span of four but also
correctly repeated all complex numbers up to three digits (comprehension problems hampering his performance only with longer numbers). He could read (in
Arabic and alphabetic codes) complex numbers up to eight digits (N of items:
50) without being affected by phonemic errors. His errors were limited to lexical substitutions (e.g., thirty-six → fourty-two); virtually no number “syntax”
violation was detected.
A test was built in which GBC was administered 75 words embedding a
number word. For example: “lotto” – lotto – embeds “otto” – eight, “alisei”
– tradewinds, embeds “sei” – six. GBC substituted vowels in 16/75 words,
four substitutions occurred on the embedded number words. The same number
words were also embedded in 75 non-words derived from the previous corpus
by changing one or two phonemes from the non-number part of the words. GBC
made 20/75 vowel substitutions, four of which were on the embedded number

The first series of tasks showed that GBC made phonemic errors predominantly on vowels in all oral production tasks besides spontaneous speech. Due to
his Wernicke’s aphasia, GBC made semantic errors both in repetition and picture
naming, resulting in neologisms that are difficult to interpret. His reading aloud
was relatively free from other errors. Therefore, we chose a reading rather than
a spoken task to investigate his unusual error pattern. GBC made more errors in
reading connected prose than single words. Because he was comfortable with
reading prose, despite the greater difficulty of the task, the investigation of GBC’s
phonological errors was therefore conducted by having him read a newspaper
article aloud, as this allowed us to collect a more sizeable corpus of errors.
The choice of a reading task is particularly appropriate for Italian, because
the relationship between phonology and orthography in this language is “oneto-one” and transparent to a large degree.
GBC spoke Standard Northern Italian, a system with five vocalic phonemes
/a/ /e/ /i/ /o/ /u/ (Mioni, 2001). Unlike English and many other languages,
vowels in Italian are always full. There are no phonological reduction processes that result in the production of the neutralized central vowel “schwa”.
This may explain why selective substitution errors on vowels have not been
noticed in the aphasiological literature, which has focused predominantly on
English.

2.5. Comparing vowels and consonants
GBC’s newspaper article was written in standard Italian and consisted of
2882 words (with an average of 2.8 syllables), with 15,042 letters (with 6963
vowels and 8079 consonants). He made 632 word errors (errors were counted
only in cases where the target word was still identifiable); 436 of the errors
were single phoneme errors. These errors were distributed over 410 words
because sometimes multiple phonemic errors occurred in the same word (e.g.,
“sfortanotamente* ” instead of “sfortunatamente”). Of the errors, 69/436 were
deletions (“pitre* ” instead of “pietre”) and 49/436 were insertions (e.g., “ni”
within “ultrinifindomentalista* ”, instead of “ultrafondamentalista”). Single

428

C. Semenza et al. / Neuropsychologia 45 (2007) 425–430

Table 1
Confusion matrix for single phoneme errors in full and restricted error sample (in bold)

phoneme errors were the more frequent informative type of error and thus
underwent further analysis.

2.6. Analysis of phonological errors in restricted sample
We examined only those errors that could safely be attributed to phonological factors: morphological (“potrà” – (s)he will be able to – instead of
“potrò” – I will be able to) (“diretto”—direct, instead of “diritto”—law),
and semantic (“donare”—donate, instead of “denaro”—money) errors were
excluded from further analysis. The restricted sample contained 111 non-words
(* ) substituting one (in most cases) or more phonemes in response to given
target words (“islimica* ” instead of islamica—Islamic; “ferici* ” instead of
“feroci”—ferocious; “sembre* ” instead of “sembra”—it seems). There were 407
vowels, 113 of which were substitutions. Table 1 shows the confusion matrix
both for the full sample of 410 words and for the restricted sample of 111 words.

2.7. Vowels
Substitutions were made for /a/ in 48.1%, /e/ in 22.0%, /i/ in 19.0%, /o/
in 17.8%, and /u/ in 31.7% of their respective occurrences. The direction of

substitutions was the following:
• /a/ became /e/ in 24.5% of substitutions, /i/ in 11.3%, /o/ in 58.5%, and /u/ in
5.7%;
• /e/ became /a/ in 41.2%, /i/ in 35.3%, /o/ in 23.5%, and /u/ in 0.0%;
• /i/ became /a/ in 26.3%, /e/ in 57.9%, /o/ in 10.5%, and /u/ in 5.3%;
• /o/ became /a/ in 55.5%, /e/ in 27.8%, /i/ in 5.6%, and /u/ in 11.2%;
• /u/ became /a/ in 0.0%, /e/ in 16.7%, /i/ in 0.0%, and /o/ in 83.3%.
The relative frequency of vowels in spoken Italian (Tonelli, Panzeri, &
Fabbro, 1998) is the following: /a/: 25.3%, /e/: 28.1%, /i/: 20.1%, /o/: 20.0%, and
/u/: 6.5%. Thus, GBC’s errors were determined by neither a frequency criterion,
nor by substituting another vowel at chance.
We examined whether errors were affected by phonological assimilatory
processes, such as vowel harmony. Vowel harmony does not occur in Standard
Italian, but it does in many of the dialects (e.g., Rohlfs, 1966). The most common
type of harmony (aka, metaphony) is height harmony, where a high post-tonic
vowel triggers the raising of a stressed mid vowel. We considered whether cases
where the mid vowels /e/ and /o/ were realized as /i/ and /u/ could be due to
the presence of a high vowel in any post-tonic position. Substitutions due to

C. Semenza et al. / Neuropsychologia 45 (2007) 425–430
metaphony were extremely rare; there were only six errors where tonic /e/ was
replaced by /i/, and of those only three matched the context for metaphony
(e.g., “dicibel” instead of “decibel”). In only two cases was /o/ raised to /u/, and
neither had the right context for metaphony. The patient’s performance could
not be attributed to vowel harmony.
It was conceivable that vowel substitutions were influenced by word stress
(for this analysis, monosyllabic words were excluded). Only 10.2% of errors
on vowels were located on stressed vowels. Furthermore, Italian orthographic
conventions explicitly mark accented vowels in final position (e.g., “università”—university). GBC made only one error on an explicitly accented vowel.
It appears that in fact, word stress protected vowels from substitution.
We also looked at the role of phonological distance in substitution. As an
approximate distance measure, we considered proximity in vowel space along
the backness and height dimensions. Of GBC’s errors, 97/113 (86%) were substitutions between vowels that were contiguous either on the front/back dimension
(/e/ replaced with /i/) or the height dimension (e.g., /e/ replaced with /o/), and
resulted in a vowel that was proximal to the target vowel. GBC’s vowel substitutions appeared to be influenced to some degree by contiguity in phonological
vowel space.

2.8. Consonants
There were very few consonant substitutions, thus the meaning of any computation on GBC’s production of consonants is of limited value. However,
modelling Caramazza et al. (2000), correlations were calculated between the
sonority scale for Italian consonants, the percentage of errors for each target
consonant segment, and the percentage of consonant types produced as an error:
both correlations were non-significant. Notably, sonorous consonants (i.e., /l/
and /r/) were not affected any more than other consonants. Another analysis
compared the error rates for consonants that are distinguished by dorsal features
(high, back, and low) with those for other consonants. As in Caramazza et al.
(2000), no significant difference was found.

3. General discussion
GBC’s relative vowel deficit manifests itself in production
and is considerably larger than deficits found in previously
studied cases. GBC produced almost seven times as many substitution errors on vowels than on consonants (the vowel/consonant
error rate was approximately 3/1 in Caramazza et al., 2000, and
2/1 in Romani et al., 1996).
GBC’s articulation is flawless, suggesting that the deficit is
not located in the peripheral aspects of production, such as the
motoric implementation of the articulatory gestures. Crucially,
because GBC makes very few errors on consonants and substitutes vowels with other vowels from the Italian inventory, his
performance suggests intact phonological representation of an
abstract category vowel (V) as well as intact articulatory gestures
for Italian vowels. GBC’s problem is neither one of representation nor articulatory implementation, rather one of selection of
a vowel to “spell out” the phonological slot with the appropriate
vowel quality, or melody.
GBC’s performance is best captured in models of phonological access that assume the existence of an abstract phonological
frame with place holders for sequences of consonants and vowels (e.g., Dell, 1988). To account for the selectivity of GBC’s
deficit, we propose a representation borrowed from multi-linear
theoretical models of phonology, along the lines of McCarthy
(1981). In this model, an abstract phonological frame (or CV
skeleton) is linked to separate tiers that specify, or spell out, the
content of individual consonants and vowels. Fig. 2 shows the
phonological representation for the word “gatto”—cat.

429

Fig. 2. Phonological representation of the Italian word gatto (cat). Note that
separate levels represent vocalic and consonantal slots (the CV skeleton), which
are then filled with separate phonemic specifications for vowels and consonants,
selected from a language specific inventory.

GBC’s error free production of number words in contrast,
suggests that the output phonetic forms of number words in this
patient are retrieved as (sequences of) full basic number words
with segmental information already phonetically spelled out.
That is, without necessarily generating a phonetic code from lexical (semantic) and more abstract phonological representations.
Unlike the production of other words in the lexicon, number
words can be produced via dedicated numeral (orthographic or
Arabic)-to-spoken name transcoding algorithms that need not
contain semantic information (Cipolotti & Butterworth, 1995),
and may be modality specific (Campbell, 1995). This suggests
that all patients with phonological deficits at the word activation
level would be relatively less impaired when producing number
words than other word categories. Indeed, this appears to be the
case. Cohen et al. (1997) describe a case of phonological jargon
where the patient made phonemic errors on all words except for
number words.
In summary, this study provides evidence for a neurologically based dissociation in the production of vowels and consonants. Along with previous cases, it suggests that separate
brain systems underlie the representation of abstract phonological information such as the number and sequencing of vowel
and consonant positions (CV skeleton) and the selection mechanisms that ‘spell out” the C and V slots. The selection mechanism
for vowels is separable from the selection mechanism for consonants, as it can be selectively impaired. More research is needed
to determine what factors influence the selection mechanism,
but for GBC, vowel selection did not appear to be influenced by
phonological context. Word stress assignment protected vowels
from errors. Further investigations of cases such as GBC should
allow us to specify the neurological underpinnings of this complex production system.

Acknowledgments
This research was supported in part by a grant from the Italian
Ministry of Education and from a Marie Curie Network Grant
(NUMBRA) to the first author. Thanks to GBC for his enthusiastic collaboration.

430

C. Semenza et al. / Neuropsychologia 45 (2007) 425–430

References
Béland, R., Caplan, D., & Nespoulous, J.-L. (1990). The role of abstract phonological representations in word production: Evidence from phonemic paraphasias. Journal of Neurolinguistics, 5, 125–164.
Blumstein, S. E. (2001). Deficits of speech production and speech perception in
aphasia. In R. S. Berndt (Ed.), Handbook of neuropsychology: Vol. 3, (pp.
95–113). New York: Elsevier.
Campbell, J. I. D. (1995). Mechanisms of simple addition and multiplication: A
modified network-interference theory and simulation. Mathematical Cognition, 1, 121–164.
Canter, G. J., Trost, J. E., & Burns, M. S. (1985). Contrasting speech patterns
in apraxia of speech and phonemic paraphasia. Brain and Language, 24,
204–222.
Caramazza, A., Chialant, D., Capasso, R., & Miceli, G. (2000). Separable processing of consonants and vowels. Nature, 403, 428–430.
Cipolotti, L., & Butterworth, B. (1995). Toward a multiroute model of number
processing: Impaired number transcoding with preserved calculation skills.
Journal of Experimental Psychology: General, 24, 375–390.
Cohen, L., Verstichel, P., & Dehaene, S. (1997). Neologistic jargon sparing
number words: A category-specific phonological impairment. Cognitive
Neuropsychology, 14, 1029–1061.
Dell, G. (1988). The retrieval of phonological forms in production: Tests of
predictions from a connectionist model. Journal of Memory and Language,
23, 124–142.

Denes, G., Semenza, C., & Magno Caldognetto, E. (1999). Phonological disorders in aphasia. In G. Denes, & L. Pizzamiglio (Eds.), Handbook of
neuropsychology (pp. 195–214). London: Erlbaum.
Dressler, W. U., Tonelli, L., & Magno Caldognetto, E. (1990). Phonological paraphasias versus slips of the tongue in German and Italian. In J. L. Nespoulous,
& P. Villiard (Eds.), Morphology, phonology and aphasia (pp. 206–212).
New York: Springer-Verlag.
Luzzatti, C., Willmes, K., & De Blaser, R. (1996). Aachener aphasie test. Versione italiana (2nd ed.). Firenze (Italy): Organizzazioni Speciali.
McCarthy, J. (1981). A prosodic theory of nonconcatenative morphology. Linguistic Inquiry, 12, 373–418.
Miceli, G., Laudanna, A., & Burani, C. (1994). Batteria per l’analisi dei deficit
afasici. In I. d. P. Università Cattolica, CNR, Roma (Ed.): Associazione per
lo sviluppo delle ricerche neuopsicologiche.
Mioni, A. (2001). Elementi di fonetica. Padova: Unipress.
Nickels, L. (2001). Spoken word production. In B. Rapp (Ed.), Handbook of
cognitive neuropsychology (pp. 291–320). Philadelphia: Psychology Press.
Rohlfs, G. (1966). Grammatica storica della lingua italiana e dei suoi dialetti:
Fonetica. Torino: Einaudi.
Romani, C., Granà, A., & Semenza, C. (1996). More errors on vowels than in
consonants: An unusual case of conduction aphasia. Brain and Language,
55, 144–146.
Tonelli, L., Panzeri, M., & Fabbro, F. (1998). Un’ analisi statistica della lingua italiana parlata. Studi Italiani di Linguistica Teorica e Applicata, 3,
501–514.