Feature sets

A repository of feature sets we have discussed. The feature set discussion shifted to the feature transcription effort, and that page now has the most up-to-date feature set.

Feature set 1

Based on the vocal tract variables of articulatory phonology, and used in Livescu & Glass, HLT/NAACL-04 & ICSLP-04.

Feature set 2

If we make the assumption that LIP-LOC and LIP-OPEN are always synchronized; the 4 tongue features are always synchronized; GLOT and VEL are always synchronized; and there are no substitutions of feature values, then we can collapse each set of synchronous features into a single feature, whose state space of possible values may be much smaller than the product space of the component features. Feature set 2 is the result. This is the feature set used for the baseline end-to-end recognition experiment in ExperimentsResults.

Feature name	Values	Comments
G	0=C-VO, 1=C-VL, 2=O-VO	Combination of GLOT and VEL.
T	0=D-CR-U-M, 1=A-CL-U-N, 2=A-CL-U-M, 3=A-CR-U-M, 4=A-N-U-M, 5=A-MN-PA-N, 6=A-MN-PA-MN, 7=A-M-PA-M, 8=A-M-U-M, 9=A-W-V-M, 10=P-CR-PA-MN, 11=P-M-U-MN, 12=P-W-V-CL, 13=P-W-V-CR, 14=P-W-V-N, 15=P-W-U-N, 16=P-W-U-MN, 17=P-W-PH-MN, 18=R-N-U-M	Combination of TT-LOC, TT-OPEN, TB-LOC, TB-OPEN. Values are concatenated abbreviations of those four feature values, respectively.
L	0=P-N, 1=P-W, 2=L-CL, 3=L-CR, 4=L-W, 5=D-CR	Combination of LIP-LOC and LIP-OPEN. Values are concatenated abbreviations of those two features' values; e.g. L=D-CR is equivalent to LIP-LOC=DEN, LIP-OPEN=CRIT.

Feature set 3

Based on Chang, Wester, & Greenberg, Sp. Comm. 47:290-311, 2005 (see Papers). This is the type of feature set we might use for acoustic modeling.

One drawback of this feature set as used in Chang et al. is that diphthongs get a single feature vector; e.g. [ay] is low front, although the ending articulation is presumably high. Also, static/dynamic refers to entire segments rather than to instantaneous articulations, so it probably doesn't make sense for us to use it. Not clear whether tense/lax also has this issue.

Mapping from feature set 1 to feature set 3

If we are to use different feature sets for pronunciation modeling and acoustic modeling, then we need a mapping from the former to the latter feature sets. Here's a mapping from feature set 1 to 3:

Feature set 4

A modification of feature set 3, which I am suggesting for observation modeling as an alternative that fixes some drawbacks I see with FS3. -- KarenLivescu - 31 Dec 2005

Questions:

• What should [hh] be? In FS1, it is essentially a voiceless [ah]. That's OK for some [hh]s, but in words like "human" it's more like a palatal fricative.
• Should there be a dental place other than inter-dental (e.g. for dental [n]s)? This would help model things like "in the" with no [dh] but with a dental [n]. If we have such a value, there is no way to train classifiers for them from phonetic transcriptions. It would either have to happen through some EM-like training, or by bootstrapping from manual transcriptions.
• Is it OK to not have a nasal consonant manner? I.e. is it OK to lump oral stop closures and nasal stop closures in the same manner class? It makes the feature set less redundant, but is it harder to classify?

A question with an answer (just a reminder):

• Why not just use an established feature set, e.g. the one in the IPA chart or Stevens' distinctive features?
  • Because those features refer to segments, not instantaneous articulations. The IPA ones are almost fine, but still have some segment-specific properties (e.g. "aspirated stop" means the end of the stop is aspiration noise).

-- KarenLivescu - 21 Jan 2006

Phone-to-feature mapping table

For now this is a mapping from phones to feature set 1. (See a few notes below.)

A few notes:

• Feature values are indicated with their numerical labels to save space--see feature definition table for the meanings of the numerical values.
• Dynamic phones (diphthongs, stops, etc.) have been broken into two parts.
• [dn], [tn], [dcln], [tcln] are "nasalizable" versions of [d, t, dcl, tcl] (typically in post-nasal position, e.g. "winter")
• [sil] and [epi] (epenthetic silence) have been given dummy feature values for now

-- KarenLivescu - 22 Dec 2005

A first attempt at converting Buckeye phones

I started playing around with the definitions Karen made in order to see if I could get a mapping from Buckeye phones to Feature Set 4. Karen's psuedo-code inspired me to see if I could take her descriptions and compile them into a perl program to convert the table above into feature definitions. I'm running into some inconsistencies with the mapping (some of which Karen describes above).

First: Here are the rules I'm using to map between FS 1 and FS 4. They are just slightly modified versions of her rules.

Applying the following script to the ruleset above creates a mapping script:

#!/usr/bin/perl

print <<EOM;
#!/usr/bin/perl

print \" | #phn | place | manner | nasal | voicing | lip-rd | voc-ht | \\n";
while(<>) {
    chomp;
    # hack!  choose first of all symbols
    s/,[^\s]*//g;
    (\$PHN,\$IX,\$LL,\$LO,\$TTL,\$TTO,\$TBL,\$TBO,\$VEL,\$GLO)=split;

    \$place="XXX";
    \$manner="XXX";
    \$nasality="XXX";
    \$voicing="XXX";
    \$lip_rounding="XXX";
    \$voc_tongue_ht="XXX";

# BEGIN AUTOMATICALLY GENERATED RULES
EOM
    
while(<>) {
    chomp;
    s/\s+$//;
    ($feature,$value,$mapping)=split(/\s+/,$_,3);

    $mapping=~s/GLO ([<>]?=?) CL/\$GLO \1 0/g;
    $mapping=~s/GLO ([<>]?=?) CR/\$GLO \1 1/g;
    $mapping=~s/GLO ([<>]?=?) WI/\$GLO \1 2/g;


    $mapping=~s/LL ([<>]?=?) (DEN|LAB-DEN|L-D)/\$LL \1 2/g;
    $mapping=~s/LL ([<>]?=?) PRO/\$LL \1 0/g;
    $mapping=~s/LL ([<>]?=?) LAB/\$LL \1 1/g;

    $mapping=~s/LO ([<>]?=?) CL/\$LO \1 0/g;
    $mapping=~s/LO ([<>]?=?) CR/\$LO \1 1/g;
    $mapping=~s/LO ([<>]?=?) NAR/\$LO \1 2/g;
    $mapping=~s/LO ([<>]?=?) WI/\$LO \1 3/g;

    $mapping=~s/TTL ([<>]?=?) DEN/\$TTL \1 0/g;
    $mapping=~s/TTL ([<>]?=?) ALV/\$TTL \1 1/g;
    $mapping=~s/TTL ([<>]?=?) P-A/\$TTL \1 2/g;
    $mapping=~s/TTL ([<>]?=?) RET/\$TTL \1 3/g;

    $mapping=~s/TTO ([<>]?=?) CL/\$TTO \1 0/g;
    $mapping=~s/TTO ([<>]?=?) CR/\$TTO \1 1/g;
    $mapping=~s/TTO ([<>]?=?) NAR/\$TTO \1 2/g;
    $mapping=~s/TTO ([<>]?=?) M-N/\$TTO \1 3/g;
    $mapping=~s/TTO ([<>]?=?) MID/\$TTO \1 4/g;
    $mapping=~s/TTO ([<>]?=?) WI/\$TTO \1 5/g;

    $mapping=~s/TBL ([<>]?=?) PAL/\$TBL \1 0/g;
    $mapping=~s/TBL ([<>]?=?) VEL/\$TBL \1 1/g;
    $mapping=~s/TBL ([<>]?=?) UVU/\$TBL \1 2/g;
    $mapping=~s/TBL ([<>]?=?) (PHA|PHAR)/\$TBL \1 3/g;

    $mapping=~s/TBO ([<>]?=?) CL/\$TBO \1 0/g;
    $mapping=~s/TBO ([<>]?=?) CR/\$TBO \1 1/g;
    $mapping=~s/TBO ([<>]?=?) NAR/\$TBO \1 2/g;
    $mapping=~s/TBO ([<>]?=?) M-N/\$TBO \1 3/g;
    $mapping=~s/TBO ([<>]?=?) MID/\$TBO \1 4/g;
    $mapping=~s/TBO ([<>]?=?) WI/\$TBO \1 5/g;

    $mapping=~s/VEL ([<>]?=?) CL/\$VEL \1 0/g;
    $mapping=~s/VEL ([<>]?=?) OP/\$VEL \1 1/g;


    $mapping=~s/ =\s+\"/ eq \"/g;
    $mapping=~s/ = / == /g;
    $mapping=~s/ OR / \|\| /g;
    $mapping=~s/ AND / \&\& /g;

    printf("\$%s=\"%s\" if ((\$%s eq \"XXX\") && (%s));\n",$feature,$value,$feat
ure,$mapping);
}

print <<EOM2;

# END AUTOMATICALLY GENERATED RULES

    print join(" | ","",\$PHN,\$place,\$manner,\$nasality,\$voicing,\$lip_roundi
ng,\$voc_tongue_ht,""),"\\n";
}

EOM2

However, when I use the script, some of the definitions I get are not sensical. So -- an opinion question -- should I try to get the rules right or just fix the FS4 definitions by hand?

Here's the result:

-- EricFoslerLussier - 02 Jan 2006

I will work on the rules. The main nonsensical ones seem to be the vowels. They seem to be very hard to convert between articulatory phonology-style features and IPA-style features. In the meantime, here's a fixed table for FS4 (a few notes below):

Notes:

• axr = GLI?
• We will need more front-back/low-high values if we want to distinguish all of the vowels (ah vs. ax, uw vs. uh).
• Some other phone sets are indistinguishable (e.g. t vs. s). Some are perhaps OK left undistinguishable (e.g., r vs. axr vs. er, uh vs. ow2, ih vs. ay2).

-- KarenLivescu - 17 Jan 2006