Report: Acoustic cues to allophony

The National Science Foundation grant number 0843959 “The Effects of Acoustic-Category Definition on the Learning of Phonological Classes” allowed us to investigate one aspect of the input that can help infants determine allophones, sounds that are present in the ambient language, but not lexically contrastive. In previous work, we had found that English learners stop attending to the nasal/oral vowel contrast (which is allophonic in English) between the ages of 4 and 11 months, likely too early to be aided by lexical contrasts. Therefore, we hypothesized that there may be low-level, acoustic cues that could bias infants' attention away from allophones.

To test this hypothesis, Amanda Seidl and Kris Onishi gathered a corpus of infant- and adult-directed speech from moms living in the American Midwest and Montreal (respectively). The moms were given toys whose names contained tense and lax vowels (phonemic in English, allophonic in Quebecois French) and nasal and oral vowels (phonemic in French, allophonic in English); as well as the point vowels /i,a,u/ frequently measured in infant-directed speech. This page aims at making publicly accessible a host of results, primarily from the methodological end. Feel free to ask me (alecristia gmail.com) for more information, and let me know if there are better ways of doing these analyses.

Our scripts were developed in Praat and R so that the scientific community could benefit from them with minimal investment (Praat and R are free; and these scripts only assume the user can detect vowel onset & offset using the instructions for vowel tagging given below, so they can be used by non-experts). Admittedly, R is not very user friendly, but the scripts below should work with minimum adaptation. Finally, given our interest on what babies might be able to do with the acoustic correlates, we favored perceptually-based parameters (over e.g., 39 parameters, which are doubtless more powerful, but less specific -- comparison with them is ongoing).

*Note: All scripts assume a certain structure of data, so they might not work unless you follow it too. Most crucially, all of our sound files are WAV (not wav, or aiff), so if you use a different extension, that's something that should be changed throughout.

Human tagging

script for finding the target words or syllables and videos with instructions in French [many thanks to Inga Vendelin] and (Argentinean-accented) English

script for coding the vowel onset and offset within tagged words/syllables and a pdf with instructions and examples

Annotated bibliography on tenseness, nasality, and Québécois phonology (with an emphasis on the loi de position/closed syllable laxing)

Frequency of occurrence of tense/lax and nasal/oral vowels in French and English: zip containing scripts and summary of results

Documentation of the acoustic-phonetic analyses' tests

Scripts:

Acoustic analyses: the analyses proceed in three steps, using Praat and R. For the time being, no instructions are given, and there are few annotations in these acoustic analyses scripts. (For a simpler version of this script, visit this page)

Phonological status across English and French: data, scripts for main and supplementary analyses.

Phonological status across IDS and ADS in English: data, script for main and supplementary analyses.

Comparison registers across prosodic positions (not on allophony): data, scripts for all analyses.

Comparison registers across monolingual/bilingual status (not on allophony): data, main analyses, supplementary analyses.

Results:

Phonetic instantiation of allophonic versus phonemic dimensions in English versus French, in monolinguals and bilinguals (with Amanda Seidl, Kris Onishi, & Golnoush Alamian)

Some plots describing the English and French data.

Supplementary analyses on the English and French comparison and alternative pipelines.

Draft of the reference (published in Journal of Phonetics in 2014)

See also some cool reanalyses using MFCCs.

IDS-ADS comparison (within English, the French ADS was not available; with Amanda Seidl)

IMPORTANT: While archiving this study, I found errors in the analysis pipeline that made some statements in the original article factually wrong. Please read a description here. These have been corrected in the final version of the article. All of the following have also been corrected.

Some plots describing the IDS & ADS data.

F1xF2 scatterplots by vowel category & register (used to set cutoffs for acceptable F1 and F2 values) at 40% and 80% of the vowel duration.

Distances as a function of the minimum number of vowels

Supplementary analyses on the IDS-ADS comparison (specifically a MANOVA on untransformed correlates; and analyses separated by acoustic correlate).

Vowel space plot in color.

Reference (published in Journal of Child Language on 2014).

Are IDS-ADS differences modulated by prosodic salience (strong versus weak vowels, words uttered sentence-medially versus sentence-finally)? (with Yuanyuan Wang & Amanda Seidl)

Plots relevant to all results (including distributions of the variables before and after transformations).

Draft of the reference (published in Journal of Child Language 2015). See also a chapter attempting to draw generalizations across various prosodic classes.

Comparison of inter-language diversity in in monolingual versus bilingual speech (with Kyle Danielson, Amanda Seidl, Kris Onishi, & Golnoush Alamian)

Supplementary tables including effect size and power calculations.

Reference (published in Journal of the Acoustical Society of America in 2013).