Analysing Sound Patterns [draft]

Introduction

This project aims to produce a free tool to assist literature students and academics in analysing sound patterns in texts. One project ended in May 2004 and the other ended in May 2006.
The ASP project (part of an M.A.) closed in May 2004. The source code is available. If you have java you should be able to download ASP.jar and type "java -jar ASP.jar" to run the program. Works for me on Mac and Unix machines (though you may need to widen the window sometimes). I've not tried it on Windows, though it was developed there.
The SPRIT project (part of an M.A.) closed in May 2007. The source code is available. See the README.TXT file in the dist folder for details.
The source code of a matlab prototype of the SPRIT project is available. Download the code and run main. The prototype doesn't analyse the text, but once the rhyme and stress patterns are described, it can identify forms. See the documentation

Sound effects have been used extensively in poetry. In oral poetry these effects assist in remembering the poem. The effects also have an aesthetic function working at the sound level alone. Sound can also interact with meaning in several ways: onomatopeia; emphasising certain words; providing structural support by binding together spatially distant words, etc. More generally, foregrounding the language is a way to produce poetic heightening, differentiating the text from prose.

Identification of regular rhymes is uncontroversial. The strict sound patterns of Welsh poetry should be easy to detect. In Anglo-saxon alliterative verse the effects are somewhat less regular. More subjective is metrical analysis - identifying stressed and unstressed sounds then organising them into units. More subjective still is the detection and interpretation of the effects in free verse.

The proposed program may be able to identify sound of the effects and for issues it can't resolve, it could provide objective evidence to aid interpretation by humans. Patterns can be typographic, conceptual, syntactic as well as sonic. Given the connection between typographic and sonic patterning and the ease of identifying typographic patterning, some typographic features will be analysed too.

Sound Effects

There are many types of sound effects to detect. Here are just a few examples to illustrate the variety In metrical poetry line length is determined by sonic features

What the critics say

How do readers respond to these effects, and what interpretations do they make? Here is an extract from Ruth Padel
Achilles, the key name, appears in every stanza. Its central syllable is repeated in the first stanza ("until", "filled", "building", with a sideways echo in "curled" ...), reappears in the second, resonates in the third with "built" and "still" (plus an echo in "full"). and reaches a climax in "killer": bringing out the fact that "Achilles" has the sound of that word "kill" in his name
Here's something by Tony Curtis
I can hear the insistence of an 'ight' sound in the words 'night...quietly...late...thought...late...silent' [p.148]
This is from Tomlinson's "The Atlantic" 
   Then curded, shallow, heavy
With clutering bubbles, it nears
   In a slow sheet that must climb
Relinquishing its power, upward
   Across tilted sand.
It's described by Ian Brinton in Tears in the Fence (No.49) as follows
The evocation is of a bulk of water 'curded' and 'heavy' which is impelled forward by the speed of 'nears' at the end of the line. The movement is then counteracted by the 'slow sheet' and this process of slowing down is aided by the lack of a comma until, when one does come, we have been almost ground to a halt with 'upward'
(one could also add that 'tilted sand' has the sharper, hissier sounds of water draining through sand)

Here's part of a review by Forrest Gander of Jorie Graham's "The Scanning" (Boston Book Review, Summer 1997)
We hear first the echo of "kiss" in "its" and "mathematics". But even before those three notes are reinforced by "hiss", "missed," "distance," and "pianissimo," Graham introduces a counterpoint, the growling consonance of "glint," "gripped" and "glides" and the long o's of "show" and "over". Look how the word "show" recollects the second syllable of "harrowing" from the second line, and prepares our ears for the deep vowels in "pianissimo," "telephone,"
and Eliot's 
    Arms that are braceleted and white and bare 
    (But in the lamplight, downed with light brown hair!)
is described thus by David Trotter ("T.S. Eliot and Cinema", Modernism/Modernity 13.2 (2006))
The intensity of Prufrock's arousal produces or is produced by an intensification in the verse. By comparison with its sparse and evenly paced predecessor ("and white and bare"), the line describing the hair on the women's arms seems positively swollen: the echo of "lamplight" in "light brown hair" and the internal rhyme on "downed" and "brown" fill it from within with the sameness of sound, with emphasis (p.243)

This kind of criticism raises various issues

  1. whether all the perceived patterns exist - they may be the result of selective highlighting in the text. "Lit. crit. has a very bad record for selective quotation and selectively quoting supporting evidence while excluding all contrary data points." [J.C.]
  2. if these patterns exist, are they accidental (i.e. are they as likely to occur in non-literary language)? - texts (particularly literary ones) will have bunched patterns of sounds. For example
Even if the program can't help resolve these issues it should be able to supply objective evidence. Moreover it should be able to Questions that the program could help answer include

The program won't be able to resolve any issues relating to meaning or emotional effect. The following examples are beyond the program's scope

The best the program can do in such situations is to present the sonic evidence impartially to facilitate human interpretation. For example, in the Auden poem above, maybe other lines have at least as many fricatives.

Implementation

The program will be written using Java. Ideally it will be extendable so extra modules/add-ons can be plugged in. The project might be broken down into these stages
structure
Each processing stage will produce a file in a standard format (if an appropriate one exists) that won't normally be viewed by users, though it may be advantageous to make them tweakable by people who know what they're doing, so where possible they'll be text files. If the output of the text-to-phoneme stage was editable, then users could do fine tuning themselves. For example the rules on whether one pronounces "the" as "thee" or "thu" aren't clear, and may require intervention.

Issues

Computing Appendix

Literary Appendix

Why is this hard?

Let's first take a simple case. You wouldn't think that identifying a poem as a sonnet is difficult - after all, schoolchildren do it. You've probably done it yourself. Before looking at the meaning, one might count the lines and the syllables, then look at the rhythm and rhyme.

Even a fixed form like a sonnet has varieties with family resemblances. Much of this analysis depends on assessing deviations from some model form, but even with a form as venerable as a sonnet, finding the most appropriate model form is non-trivial.

There are 100s of other fixed forms (increasingly being imported and adapted from other cultures). Identifying the form of a poem may require searching for features in a way that isn't useful when studying a poem in a known genre (similarly, the factors used to find a face amongst a crowd of objects may not be the factors used to distinguish between one face and another).

Some fixed forms are hard for readers to detect (though might be easier for computers). E.g.

Nowadays poetry is predominantly free-form, and that's where the trouble really begins. Patterns of sound aren't regular - indeed their existence shouldn't even be assumed. Here is an extract from Ruth Padel's book showing the scale of the problem

Achilles, the key name, appears in every stanza. Its central syllable is repeated in the first stanza ("until", "filled", "building", with a sideways echo in "curled" ...), reappears in the second, resonates in the third with "built" and "still" (plus an echo in "full"). and reaches a climax in "killer": bringing out the fact that "Achilles" has the sound of that word "kill" in his name

Why bother analysing in this way? Well, though texts have meaning as the result of matching words to the world, there's also a non-representational aspect. Such effects have been studied in music, where expectation, variation and repetition produce something which at least some theorists term "meaning". These non-representational effects can be foregrounded by reducing the literal meaning of the text and increasing the density of sound effects. Charms and incantations exploit sound in this way.

Although a computer might not be able to detect all the features of interest to humans, I think it should help add some much-needed objectivity to discussions. In particular, the ability to perform bulk processing makes possible investigations that can't be done by hand.

Why bother?

"Take care of the sense and the sounds will look after themselves" (the Duchess to Alice)

From the start of this project we've made it clear that the meanings of words won't be determined. Words are going to be atomised into phonemes (seemingly destroying any hope of recovering meaning) and then patterns sought.

People have always known that some words have sounds that are related to their meaning (chickens cluck and cows moo) but when Saussure announced the arbitrariness of the sign and emphasised that the words in a language are meaningful by virtue of their difference from each other, the relationship of sound to the world was overlooked.

Several changes in recent years have rekindled interest in sound

All this is leading people to believe that "Arbitrary and conventional is a fitting description of distinctive sounds, less so of words, even less of sentences, and beyond that scarcely fits at all", "Language. The Loaded Weapon", Bolinger, Longman, 1980, p.18.

The rise of the status of sound (and of pragmatics) helps especially when considering Wallace Stevens (increasingly seen as the major 20th century US poet).

So rather than merely supporting the meaning, sound can be a creator of meaning. Speech recognition is performed in a different place in the brain to where other sounds are heard, and music activates several widely spaced brain regions. Tsur et al suggest that some poetic effects are caused by this contribution from the sound/music processing regions when processing words. But current methods to identify these effects (performed by humans or computers) aren't reliable. Although some sound pattens have been analysed for centuries, many of those used in free-form poetry haven't been systematically studied, in particular the "Divergent sound patterns [that] form interwoven threads, move crisscross over relatively large areas of text, and are diffused in an unpredictable order. These sound patterns tend to be perceived more or less unawares and to fuse in undifferentiated but intense musical texture characterized by an emotional quality" [Tsur p.57]

A program that could do some of the analysis would be a valuable and timely contribution.

Objections

Basil Bunting is quoted (probably incorrectly) as saying that "Poetry is only sounds". In "Bunting: the shaping of his verse" (Peter Makin, 1992, OUP), Makin looks at some of the attacks made on this extreme position. William Empson (in "Seven Types of Ambiguity") and Peter Dale (in "Bunting and the Quonk and Groggle School", Agenda 16/1, 1978, p.55) both put forward objections addressed by Makin in a section entitled "On the Centrality of Sound in Poetry" (p.337). According to Makin (p.247) objectors often

Makin thinks that Bunting meant that sound patterns were only essential in poetry, and that poetry could exist on sound-play alone (Bunting said "It is perfectly possible to delight an audience by reading poetry of sufficient quality in a language it does not know", ("The Poet's Point of View", Diary of the North-Eastern Association of the Arts, 1966)). He thinks that "Bunting's main claim ... is not that the sound will tell the hearer the meaning...It is that the sound gives us something as complex, as internally differentiated, and therefore as structurable (by the maker) as that semantic map, with all its finesses." and that "What we analogize from, in poem-sounds, may not be the sounds of consonants and vowels themselves, but of the shift from one type to another; or from one movement of words to another" (p.246). This, I think, was also Valéry's position - sound and sense can be consciously separated variables.

References

  • More notes

    • 2006-7 Additions

    I've been trying to think of the factors that determine the impact of a sound. As a rule of thumb I've decided that a sound is strengthened if

    I've knocked together something in matlab which calculates and displays these figures. It's more psychologically realistic than merely counting the sounds.

    Here's some functionality which could be contained in add-ons or could pro tem go into the core

    Other sources

    We're grateful to the following for their help May 2011
    Tim Love mail