Steve Evans Notes on “HiPSTAS Poetry Tagging” Project (First Trial of 1K Files) January 31, 2014 Overview: What follows is a sort of “diary” of the work I did on the tagging project following our Virtual Meeting of Jan. 31, 2014. Most of that work was preparatory, but I felt good about the little bit of actual tagging I eventually got to. Section 5 contains some “spoilers,” so hold off reading that section until you’ve had time to form your own impressions (or read it now and trust yourself to forget all about it before you do your own tagging). §1— Problem: In following the Guidelines for “Logging In to Poetry Tagging Test,” I found that I was not seeing the correct options for Item 3. Solution: I wrote to Tanya and Tony and learned that my username had not yet been associated with the project. Tony took care of that (within a few minutes of my initial query). §2— Action: Reviewed “Overview of Poetry Tagging Page” and “Tagging Page” sections of Guidelines. (6 minutes) Action: Reviewed “8 Transcriptions of Speech – The TEI Guidelines” (http://www.teic.org/release/doc/tei-p5-doc/en/html/TS.html). (30 minutes—but felt like I could spend a lot more time on it.) Notes: I’ve not spent time with the TEI system before now and there is a lot that I would like to go back over more slowly and/or with more guidance from someone well trained in the protocols. The terms that we adapted for the Tagging Exercise are clustered at the end of 8.3.6 (on “Shifts”). §3— Action: Reviewed Guideline notes for six tagging categories. Notes: The first five are conceived of as spectrums: • Tempo or speed: Slow to fast • Rhythm: (Highly, notably) rhythmic (e.g. metrical, hence “beatable”) to nonrhythmic • Loudness or intensity: Soft to loud

Pitch: Low to high (decision point: will you categorize relative to gender indicators—tacitly or reflexively—or will you attempt to assign tags independent of gender indicators) • Tension or articulation: Indistinct to distinct The sixth category (“voice”) is a check-all-that-apply opportunity. The acronym TRIP-TV might be helpful here. I’m pretty sure the TRIP components are all quantifiable by ARLO (and most other sound analysis platforms). •

§4— Action: Navigated to ARLO “Overview of Poetry Tagging” page. Received first tagging opportunity and reviewed the interface page (shown as a screen save provided by Eric during today’s virtual meeting). Notes: I first went over using my Safari browser, then saw that the “Half-Speed” functionality is Chrome-only, so switched (and re-logged in, etc.). I noted the fact that Tempo, Loudness, and Pitch all have optional “Change” or “Range” corollaries, and “Tension” has a “(degree of) Articulation” corollary. Reviewed gray bar to left of screen, with “Silence,” “Chatter,” “Applause,” “Laughter,” “Music,” and “Not Rateable” options. As I understand it (confirmed by today’s conversation): If the clip is “a good example of” one of the five phenomena named, one marks it as such AND SKIPS IT. If it is not suited to serve as a good example of the phenomenon, one uses “not rateable” and SKIPS IT. •

So the left-hand gray box captures impressions about the CLIP as a whole: what is it an example of? (The prior and governing judgment might be phrased as: “this doesn’t have the kind of material we’re looking for, BUT it IS a GOOD EXAMPLE of X”).

•

The central columns capture impressions about the voiced UTTERANCE in the clip.

•

The right-hand column is meant to capture impressions about the VOICE making the utterance plus—asymmetrically—an impression about genre (is this an example of poetry or “not poetry”).

As to the visualization of the two-second clip, I note that our frequency range is between 60Hz and 12,000Hz. Typical/normative adult male range (for fundamental frequency) is 85-180hz, typical/normative adult female 165-255hz. Standard telephony range is: 04000Hz (voice channel), 300-3000Hz (voice bandwidth). A visualization from the web (that would perhaps be more helpful if rotated 90 degrees)

S. Evans Tagging Trial Log 2

§5— Action: Actually played first two-second clip. I only played it once, went immediately over and clicked “not poetry,” then filled in the categories moving right-to-left. Perceived gender: Masculine. Already I notice that “Rhythm” poses a question for me. If I detect an ordinary speech rhythm, I think I have to mark “arrhythmic” because that actually means not metrical at all. Had to double check TEI glossary: “Staccato” in TEI is glossed as “every stressed syllable being doubly stressed” while “legato” is “every syllable receiving more or less equal stress.” Also, “scandent” is defined as a variant of ascending pitch, where “each succeeding syllable [is] higher than the last, generally ending in a falling tone.” Action: Played second two-second clip (with headphones this time). Perceived gender: Masculine. Again was drawn to the vocal characteristics, which I marked first, then proceeded right-to-left. Decent confidence in the various judgments. No replay. Third clip. No strong impression of vocal quality. Perceived gender: Masculine. Clicked “staccato” but a little uncertain about it. (I notice that I pause before hitting “submit” because the next clip cues up automatically and I want to be ready for it.) Fourth clip. No lexical content, perhaps some technological timbre. Marked not rateable. Fifth clip. Recognized (masculine) speaker’s voice and the “genre” (announcement of a sponsoring institution and/or broadcasting channel; upon reflection, letter names, so probably an URL being read aloud). ID: Al Filreis. Began by marking “not poetry.” Selected “legato” because each letter-name was given roughly equal stress.

S. Evans Tagging Trial Log 3

Sixth clip. Multiple (masculine) voices speaking language(s) other than English. (I’m closing my eyes to listen, then clicking categories without consulting the visualization. Perhaps I should linger for a moment to study the visualization before accepting next clip for tagging?) Seventh clip. Impression of hearing end of one phrase and beginning of next. Perceived gender: M. 8th clip. Complete phrase: “there’s no answer to it.” ID: Steve Benson. Perceived gender: M. Had neglected to assign a LOUDNESS indicator so heard the clip again. Slight rising movement (not borne out by visualization). 9th clip. Perceived gender: M. Assigned pitch relative to gender. Sort of worked up, energetic (even hyper). 10th clip. Perceived gender: M. Hiss of sibilant, then phrase concluding “memento mori.” British accent. Did a second listen, then a third at half-speed and changed RHYTHM indicator to “moderate” on the basis of stress distribution. 11th clip. “…essentialism … to … to say.” Perceived gender: M. Sense of a slightly stammering delivery (duplicated of “to…to”). Visualization shows maybe eleven vertical bands. I marked most of these with the “umarked” (moderate, unremarkable) term. 12th clip. Perceived gender: M. Technological timbre (something resonating). Easy to mark “not poetry.” One of the male voices I might mark “less resonant” if that were an option (spectrum: “thin—thick,” probably corresponding to condition and disposition of vocal cords). §6— Action. Returned to “Overview of Poetry Tagging.” Selected “Visualize All Tag Classifications.” On the pull-down PLOT menu, should read . Plotted “pitch” and looked at the results for TAG ID 3561126-38, assuming those to be the first twelve that I’d just marked. If my surmise that the first twelve clips contained eleven male voices (and one without vocal content), then the interesting category is probably “high” (green band). So 131 and 138, primarily, and 134 and 136 also. §7— Action. Finished up and circulated these notes, for what they’re worth.

S. Evans Tagging Trial Log 4

Steve Evans Notes on “HiPSTAS Poetry Tagging” Project (First Trial of 1K Files) Installment 2, Sections 8-14 February 1-4, 2014 8—MORE ON LEGATO and STACCATO (2/1/2014) 8.1 Thinking more about LEGATO and STACCATO this morning. These involve deliberate (and notable) equalization of STRESSES, in the former case by deleting or promoting unstressed (-s) syllables, in the latter by increasing the stressed (+s) syllables to double strength. So, likely to be pretty rare among the samples. 8.2 I had initially thought one might mark LEGATO any segment that exhibited a certain uniformity of pitch (perhaps connected to Marit’s explorations of “monotony”). But that, I see now, would be a misapplication of the marking. Using the “narrow” option under PITCH RANGE is the better option, I think. 9—CATEGORIES FOR A SUPPLEMENTAL SPREADSHEET (In Progress) (2/2/2014) 9.1 I initially wanted to create a supplemental spreadsheet (1) to track (the controversial) “perceived gender” category (Male, Female, Ambiguous); (2) to mark clips in which the voice is recognizable to me; and (3) to note what sections of a phrasal unit seem to me to be present in a given 2-second clip (beginning/initial, middle/medial, end/terminal). 9.2 I also added check-mark categories for TEI constituents VOCAL, KINESIC, INCIDENTAL (8.3) and for WRITING (8.3.4). 9.3 Comment: The danger of course is that one “overloads” the marking that needs to be done for each clip and slows the process down. If I find that the additional categories are unwieldy, I’ll drop them. 10—SECOND TAGGING SESSION ARLO TAG EXAMPLE IDs 3561138 -3561180; 43 clips in approximately 2 hours (2/2/2014) 10.1

Spreadsheet Notes: 10.1.1 The three TEI Constituents (8.3) proved to be pretty gratuitous in this session. I only checked one of them (INCIDENT) once (for the fascinating clip 3561-179). Next session I’ll move them over to far right (less utilized) side of the spreadsheet and/or abandon them. 10.1.2 I couldn’t keep myself from adding a column for what I called a “onelisten TRANSCRIPT.” After listening to the clip, via earphones, with my eyes

S. Evans Tagging Trial Log 5

closed, I would transcribe what the lexical content of what I thought I had heard or note the obstacles to rendering such a lexical transcript. 10.1.3 I also ended up adding a checkbox for TECHNOLOGICAL TIMBRE— perhaps ARTIFACT would be the better term—for those clips that exhibit (are good examples of) sounds one associates with a given recording or playback device. Clips that one might do well to consult with Michael Hennessey or Steve McLaughlin about. Note: Gloss on “Sonic Artifact”: http://en.wikipedia.org/wiki/Sonic_artifact 10.1.4 I added a column for notes on the ARLO VISUALIZATION but couldn’t come up with a consistent way of glossing what I was seeing. I tried to count the major vertical “bands” but then didn’t know what to do with the more continuous sounds (which create horizontal bands). 10.1.5 I also added a column for ACCENT. I marked 146 “British” and 175 “U.S. Southern.” This is perhaps even more problematic than PERCEIVED GENDER since accent (as a perceived divergence from an unmarked norm) is subject to lots of situational biases. 10.1.6 I did not create a separate column for NOT ENGLISH (a feature Hartwell suggested during our January virtual meeting), but I did note that fact that clips 140, 155, and 171 were not English. 10.1.7 Clips 147, 164, and 165 seemed like good examples of SINGING to me. But we lack an official ARLO TAG for that at this point. 10.1.8 In terms of IDENTIFICATIONS, I had confidence approaching certainty for only six clips of the forty-three (13.95%) I tagged in this session (clips 142, 160, 169, 170, 175, and 180). The clips featured three cis-female poets who identify as straight and three cis-males, one of whom identified as gay. I was a little surprised that I could identify 175; the others struck me as “easy” (though note that 142 consists of a fragment of a single syllable). The voices of two speakers (both dead) are known to me only through recordings, the other four are familiar through a mixture of “live” and recorded interactions. 10.1.9 Perceived gender results: M (30 clips), F (10), N/A (3). 10.2 Question: In the optional ARTICULATION box (below the obligatory TENSION box), the options include “none.” Similarly, under TEMPO CHANGE there is a button for “same tempo.” In both cases, I wonder what is the difference between “N/A” and “none” or “same tempo”?

S. Evans Tagging Trial Log 6

10.3 Suggestion: I do find myself wishing we’d adopted a five-choice menu for “Pitch,” with intermediate steps between “Low” and “Middle” and between “Middle” and “High.” 11—THIRD TAGGING SESSION; ARLO TAG EXAMPLE IDs 3561181 -3561225; 45 clips in approximately 2 hours (2/3/2014) 11.1 Action: Revised supplemental spreadsheet (D2) in accordance with remarks above (§10; cf. §13 for headings). Added a column for NOT ENGLISH. 11.2 Action: Tagged and marked supplemental spreadsheet for 20 clips (approximately 48 minutes). 11.3 Comment: I am interested in the variety of the VISUALIZATION PANEL returns, which continue to defy my desire to annotate what I’m seeing. (There’s something of a “Rorschach test” feel to it.) The way the energy clusters, the gaps, the distribution patterns, all are beginning to correlate, for me, with what I’m hearing. But I feel like I’m still early in the learning process. 11.4 Action: Tagged and marked supp. spreadsheet for 25 additional clips (approx. 48 minutes). 11.5

Result: Perceived Gender (update) F 19 clips of 88 M 59 clips I (indeterminate) 00 I/M (indeterminate shading M): 02 I/F (indeterminate shading F): 00 n/a 08* *=of which 1 possible M and 2 possible F (plural)

12—(2/4/2014) 12.1 Comment: Either I’m bad at detecting the Change/Range phenomena we have grouped as optional fields under TEMPO, LOUDNESS, and PITCH, or the two-second increment just makes those difficult to perceive. I seem to be able to establish a base line (or to form an overall impression, a sort of sonic gestalt) but I lack the ability to detect variations from that base line within so brief a window. 12.2 Comment: One of the more humorous things I’ve found myself doing while tagging is repeating aloud, often more than once, the vocal content of the clip, in a kind of PLAY IT / SAY IT loop that aids my introspection about what I’ve heard. I suspect that if I recorded these attempts at mimicry, it would be possible to show that I was aware of sonic features—like changes in tempo, loudness, and pitch—that I was not able to make explicit enough to consciously and confidently tag.

S. Evans Tagging Trial Log 7

13—Column Headings for Supplement Spreadsheet, 2nd Draft (2/4/2014) A B C-E F G H I J K L M-O

Arlo Clip # Perceived Gender Phrasal Unit (TEI=utterance): initial, medial, terminal Writing Identification ~English (not English) Sonic Artifact (or, technological timbre) Single-listen transcript Visualization Panel notes Accent, or vocal quality comment TEI Constituents: Vocal, Kinesis, Incident

14—Stray Remarks on Clips 181-225 14.1 Clip 213 is a good example of STATIC (the Red, Green, Blue banding at the top of the visualization is especially vivid.) 14.2 Clip 214 could serve as a good example of a PAGE TURN, with nice spacing on either side. But I’d like some confirmation that others are heading it as I do. 14.3 Clips 200 and 220 were relatively indeterminate for me in terms of “hearing gender.” 14.4 Clip 187 ends with audible breath of the kind TEI would label VOCAL (“anthropophonic” but not LEXICAL). My hypothesis is that this element would be overrepresented in other files by the same author (Robert Creeley), or might even serve as a finding aid for further files by that author. 14.5 Clip 212, recognizable as one of Jackson Mac Low’s “phoneme dances” for multiple voices, exhibits some interesting downward-sloping shapes in the visualization. 14.6 The voice in clip 206 struck me as “strident.” But by “voice” I probably really mean “manner of delivery.” 14.7 The voice in 211 struck me as “gurgly.” In this case, it really is a vocal quality (that one would expect to be audible in most utterances by the speaker) that I’m pointing to. I’m curious about the correlation between that characterization and the gaps in the fifth to seventh horizontal intervals in the visualization. 14.8 IDENTIFICATIONS: For this batch of 45 Clips, I could confidently supply an identification in only 7 instances (15.5%). All but one were men.

S. Evans Tagging Trial Log 8

14.9 Clip 202 is a little inscrutable, but I think I’m hearing the button on a recording device being pressed and a faint voice saying “the second cassette.” Might such instances of DEVICE SELF-CAPTURE (a cassette recorder recording one of its buttons being depressed) be worth tagging? As a sort of parallel to our collection of ARTIFACTS? 14.10 On the basis of the single distinct phoneme audible to me in Clip 209 I felt confident in assigning gender to the speaker. 14.11 For clip 193, I first added to the judgment “Not English” a guess that it was French that I was hearing; on a second listen it was clearly German.

S. Evans Tagging Trial Log 9

Steve Evans Notes on “HiPSTAS Poetry Tagging” Project (First Trial of 1K Files) Installment 3, Sections 15-18 February 10-26, 2014 15— FOURTH TAGGING SESSION; ARLO TAG EXAMPLE IDs 3561226 3561250; 25 clips in approximately 54 min or 1.86m/clip (2014-02-10) 15.1 Comment: I was traveling for a few days and unable to do further work on the tagging project. As I resume, I’m interested to see whether I might train my ear better to recognize variations in pitch. 15.1.1 From the Pronuncian website : Why are English intonation and pitch so complicated? Understanding the uses of intonation is one of the most complicated aspects of spoken English for a number of reasons. • Research on intonation is difficult to conduct. Due to the differing nature of vowels and consonants, and voiced verses unvoiced consonants, even computers still have difficulty measuring human pitch accurately. • Intonation does not occur in steps, like a ladder, but rather by gradation, with one tone blending into the next. • Intonation is relative. One person's high pitch may be another person's mid-range pitch. • In English speech, intonation does not have syntactic rules governing its use. Vocabulary and grammar are relatively rigid in English, but intonation is not. • Intonation is emotional, and therefore difficult to measure. 15.2

Action: Revised for the third time the headings for the Supplemental Spreadsheet.

15.3 Comment: As I worked through this batch of twentyfive (25) clips in a little under an hour, I focused more on the “single listen transcript” (though with more double checking than that title would suggest) and less on the initial-medial-terminal markings from before. The transcript captures that information anyway. 15.4 Identifications: I could confidently identify the speaker in five clips (20%): Blackburn (227), Bernstein (228), Zukofsky (233), Olson (239), and Notley (248). But I hazarded names for five more, and had general comments on two others. 16— FIFTH TAGGING SESSION; ARLO TAG EXAMPLE IDs 3561251-3561325; 75 clips in approximately 150 min, or 2min/clip (2014-02-11) 16.1 Comment, before starting. I’ve been brushing up a bit on the phonological / phonetic account of pitch in standard textbooks. We’ll see if that helps me listen any better!

S. Evans Tagging Trial Log 10

16.2 Comment, after finishing. I plan to pause now in the hopes that at least five people will also be able to tag the first 200 clips. (Otherwise, my work is likely to get stranded.) 16.3 Comment on the two second clip. I’m impressed by the amount of information presented to the ear in second seconds, and I’m impressed by the number of impressions that can be formed in that same short interval. 16.3.1 But it is rare to find an UTTERANCE that will fit, completely, within a two-second window, so a lot of what I’m hearing is fragmentary: terminal fragments of phonemes and/or lexical units at the left margin, initial fragments at the right, and some opportunity for continuity in the middle. 16.3.2 So the impulse to try to record INITIAL-MEDIAL-TERMINAL was, perhaps, a warranted one, though it duplicates information captured in the TRANSCRIPTS I’ve been making and, even more clearly, the information deducible from the ARLO VISUALIZATION panel. 16.3.3 I wonder what will happen when we expand the interval to four (4) seconds—as I believe was discussed, and perhaps even resolved (?), at the Kelly Writers House meeting in Fall 2013. 16.3.4 Since RISING and FALLING PITCH is a transsegmental prosodic feature, not being able to orient ourselves within completed UTTERANCES probably carries as a consequence an uncertainty about which way the pitch patterns are trending. PITCH helps us mark PROMINENCE, but the latter is relative, and if we don’t have enough of a glimpse into the system of contrasts, it’s hard to pick it out. 16.4

Results for PERCEIVED GENDER. 16.4.1 For the first two hundred (200) clips, my tally of “perceived” (other systems use “probable”) gender are as follows: F I/F I/M M N/A

049 clips (27.53% of attempted IDs; 24.5% of all clips) 000 clips 004 clips (2.25% of attempted IDs; 2% of all clips) 125 clips (70.23% of attempted IDs; 62.5% of all clips) 022 clips (11% of all)

16.4.2 The “not applicable” (N/A) category overlaps largely (totally?) with the set of clips that are not “good examples” of single-speaker vocalization (applause, music, room tone, sonic artifact, static, etc.). These should show up in the results for NOT RATEABLE in the overall tagging tallies.

S. Evans Tagging Trial Log 11

17— SIXTH TAGGING SESSION; ARLO TAG EXAMPLE IDs 3561326-3561375; 50 clips in approximately 90 min, or 1.8min/clip (2014-02-17) 17.1 I thought that I’d rest after 200 clips, but then decided to make it 250 (a quarter of the total). Still concerned that the data will be “stranded” since few others seem active in the tagging process at the moment. 17.2 The stripped-down SUPPLEMENTAL SPREADSHEET now consists of an ID column to the left of the ARLO ID, then PERCEIVED GENDER, TRANSCRIPT, and ACCENT/COMMENT fields to the right. I’ve “hidden” the columns containing INITIAL-MEDIAL-TERMINAL markings, having decided that the TRANSCRIPT records that information better. I’ve retained ~ENG (not English), TT (Artifact), and VIZ PANEL columns, but probably made five marks in 50 clips. 18—Stray Notes (2014-02-26) 18.1 Perceived Gender. Looking again at the numbers for “typical” frequency ranges of men and women, I note the eighty (80) Hz gap between the low end of the male (85Hz) and the low end of the female (165Hz) ranges; the seventy-five (75) Hz gap between the top of the male high range (180Hz) and that of the female (255Hz); the fifteen (15) Hz shared by high male and low female ranges. 18.1.1 But also note that in the graphic reproduced in §4 above, which graphs vocal ranges using the piano keyboard as a guide, the “male voice” and “female voice” ranges overlap considerably (much more then they don’t). 18.1.2 Regarding vocal “plasticity” (the dialectic of “the given” and “the made” in vocalization): 18.1.2.1 YouTube clip on “FTM voice training” by “AdamTBoy,” three weeks into testosterone (T) therapy. http://www.youtube.com/watch?v=VKuVRfKO5qM 18.1.2.2 YouTube clip on “Sliding voice F2M” by “CandiFLA.” http://www.youtube.com/watch?v=VKuVRfKO5qM 18.1.2.3

There are many more of these.

18.1.3.4 2005 article on “Feminization of Voice and Communication.” http://speech-language-pathologyaudiology.advanceweb.com/Article/Feminization-of-Voice-andCommunication.aspx 18.2 “Firstness.” From the start (see §5 above), I’ve been aware of the sequence in which I become conscious of traits within a two-second clip. It might be interesting to

S. Evans Tagging Trial Log 12

have a way of tracking (invisibly to the user) the order in which information fields are completed in the tagging interface. The “loudness” of a clip might be the first thing that strikes you, and remain the most prominent attribute of the clip after you’re done. In another clip, it might be a vocal quality (right hand menu). It might happen that perceptual prominence could be mapped by the order in which we tick the boxes. And hesitancy might be indexed by “revisions.” 18.2.1 Also, how many listens did the user do before hitting “submit” (and what might that tell us about the clip and/or the user)? 18.2.2 This goes back to the sense of a “sonic gestalt” mentioned above in §12.1. 18.3 Microphone placement. In my next tagging session, I think I might experiment with a column in the SUPPLEMENTAL SPREADSHEET that correlates to the perceived microphone placement. A five-option spectrum from proximal to distal might be: Very close – close – neutral – distant – very distant. I’d like to talk more with Michael H. about this issue, but a recording made by an audience member in the fourth row (distant) will audibly differ from one made at the podium (close), or one made in a recording studio (very close). And the distal options will include more information about the composition of the room and the audience, the proximal with the speaker’s body, the presence of paper (manuscript, book, page turns), etc. 18.4 Recording quality. Another five-step spectrum might capture impressions of recording quality: Great, good, fair, poor, awful. 18.5 Listening transcripts. I will attach as a separate document the transcriptions I made of clips 138-375. 18.5.1 Key. These are not as scrupulous and systematic as maybe they would have been if I had decided on some protocols in advance, but generally, information between < > is descriptive, though sometimes I include within < > best guesses at phonemes that I’m understanding to be present but that aren’t actually clearly audible. Vertical bars | are used to make pauses, though sometimes I write or . In a few places I mark rising pitch / and falling pitch \. In some batches I try to indicate stressed syllables (+s) by using ALLCAPS (though since the transcripts aren’t in IPA, this is very approximate). Question marks within mark questionable guesses, not rising intonation patterns in the clips themselves.

S. Evans Tagging Trial Log 13