Our Lady of the Highway: A linguistic mystery

« previous post | next post »

Current text-to-speech systems are pretty good. Their output is almost always comprehensible, and often pretty natural-sounding. But there are still glitches.

This morning, Dick Margulis sent an example of one common problem: inconsistent (and often wrong) stressing of complex nominals:

We have a winding road that we drive with our Google Maps navigator on, to keep us from taking a wrong turn in the woods. We have noticed that "West Woods Road" is rendered with a few different stress patterns as we go from turn to turn, and we can't come up with a hypothesis explaining the variation. Attached is a recording. It's a few minutes long because that's how long the trip takes. The background hum is the car.

I've extracted and concatenated the 11 Google Maps instructions from the four minutes and five seconds of the attached recording:

This case is especially puzzling since the voice does very well on the first instance, and then screws up pretty consistently thereafter.

But in general, it's not trivial to guess the correct accentuation of patterns like "ADJ NOUN road". If it's a three-word name, as in this case, then the system's first version is the right choice:

But if with a different interpretation of the same three words, a human speaker might do something like the system's later versions:

At least, compare this clip from a sociolinguistic interview:

And the stress pattern of "Woods Road" would be more plausible if it were "Woods Street", since the English language has irrationally decided to assign primary stress to X in X Street, as opposed to what we do for Road, Avenue, Boulevard, Way, Alley, etc. Though the apparently phrase-final lengthening and pitch fall on "west" remains puzzling.

We could try make up a story about robot themes and rhemes gone wrong, but most likely this is just one of those weird and inexplicable things that "deep learning" systems sometimes do.

Some Googlers of my acquaintance are well aware of problems like this one — so maybe Dick's navigational narrative will become more natural, if less interesting, after a few more updates to Google Maps.

Here are a couple of relevant (if antique) references:

Richard Sproat and Mark Liberman, "Towards Treating English Nominals Correctly", ACL 1987.
Mark Liberman and Richard Sproat, "The Stress and Structure of Modified Noun Phrases in English", in Lexical Matters, Sag and Szabolsci, Eds. 1992.
"Parsers that count", 11/25/2003
"Complex nominal of the year", 11/2/2018.



30 Comments

  1. Dick Margulis said,

    September 8, 2022 @ 12:33 pm

    On the trip we recorded, we noticed only three variations. On earlier trips, we thought we heard at least five variations. A side note: my wife and I observed the street signs at the various turns. They variously read W Woods Road, West Woods Rd, W Woods Rd, West Woods, and so forth. But we couldn't correlate the sign variations with the aural variations, so that hypothesis went out the window.

    [(myl) You might check how Google Maps labels the road in its graphical form — it didn't occur to me that the system might be reading different text strings for the different road sections, though I should have thought of that. Of course the Google Maps graphics might not always be the same as what's in the database it's constructing directions from.

    Though anyhow, it should be able to infer that the it's the same road under (perhaps) different spellings or abbreviations.]

  2. Scott said,

    September 8, 2022 @ 1:25 pm

    …so that hypothesis went out the window…

    Try driving it with the car windows closed ;)

  3. Don Monroe said,

    September 8, 2022 @ 1:34 pm

    The first pronunciation seems to be good for the road whose name is West Woods.
    I can think of two meanings that would lead to the other pronunciation.
    One would be that it's the western *section* of Woods Road (like Northwest Pennsylvania Avenue).
    But the oddly long pause between West and Woods suggest to me that it the intended meaning is the *direction*: Westbound on Woods Road.

    [(myl) Though the natural way to express that would be "Slight right to stay west on Woods Road" or "Slight right to stay on Woods Road west", not "Slight right to stay on west Woods Road"…]

    It's better than what our GPS does, which is to occasionally (whisper) or SHOUT the directions. Anyway I'm impressed with how often it gets quirky pronunciations correct.

  4. Dick Margulis said,

    September 8, 2022 @ 2:28 pm

    @Scott: Thank you for that. :D

    @myl: I just looked. On the map it is W Woods Rd along its entire length.

    @Don: Regarding quirky pronunciations, New Haven's Whalley Avenue (Rhyme's with Halley) is still coming out as Wally Avenue. However, "Oak Saint Connector" was at some point corrected to "Oak Street Connector." So that's something.

    [(myl) Ah, but was it pronounced as if to mean a Saint Connector made of oak, or a connector of Oak Saints?

    I guess the logical roadway name would be the connector named "Oak Saint". But I don't get the impression that Google Maps is applying logic.]

  5. Breffni said,

    September 8, 2022 @ 3:06 pm

    My story about why X gets stressed in X Street is that streets, more so than other types of thoroughfare, are typically found in the vicinity of lots of other streets, and the need to differentiate has led to the institutionalisation of contrastive stress.

  6. Kenny Easwaran said,

    September 8, 2022 @ 3:12 pm

    My favorite was once, a few years ago, when a GPS navigation device pronounced "Turn left on Cesar E Chavez Ave" as "turn left on cesar east chavez avenue". It knew that "E" in a street name is often pronounced "east", but not quite what circumstances are counterexamples to that generalization.

  7. Scott P. said,

    September 8, 2022 @ 3:36 pm

    Regarding quirky pronunciations, New Haven's Whalley Avenue (Rhyme's with Halley)

    But is that Halley as in Halle Berry or Halley as in Edmund Halley? :-)

  8. Dick Margulis said,

    September 8, 2022 @ 3:38 pm

    @myl: It was more like Oak, St. Connector. Sort of like Mary, Queen of Scots.

  9. Dick Margulis said,

    September 8, 2022 @ 3:42 pm

    @Kenny: I once approached an on ramp point to NY City that was rendered as "north wye city." That was in the same vicinity as the state highway sign warning "no permitted vehicles allowed."

    @Scott: as in Bill. And if that's before your time, then yes, like the astronomer.

  10. Ross Presser said,

    September 8, 2022 @ 3:53 pm

    Google Maps consistently says "West Whitman Bridge" for the Walt Whitman Bridge into Philadelphia. I guess given that there is a *lot* of signage that abbreviates it as "W Whitman Br" it can be explained …

  11. Philip Taylor said,

    September 8, 2022 @ 3:56 pm

    Erm, Dick, I'm old enough to remember Bill H. and his Comets, but I remember his surname as being "Haley", not "Halley" …

  12. Stephen Hart said,

    September 8, 2022 @ 6:33 pm

    Siri, in the Apple iPhone/CarPlay universe also occasionally gets the emphasis wrong, though I don't think I've ever heard it say the name of a single road differently.

    OTOH, I've also heard live readers on audiobooks get the emphasis wrong, as if they had never read the text before.

  13. Stephen Hart said,

    September 8, 2022 @ 6:41 pm

    Philip Taylor said "I remember his surname as being "Haley", not "Halley" …"

    Hay lee is also the way I've always heard it.

    Wikipedia has heɪli, which http://ipa-reader.xyz reads as Hay lee.
    https://en.wikipedia.org/wiki/Bill_Haley

  14. Charles in Toronto said,

    September 8, 2022 @ 6:42 pm

    Many years ago when TTC first introduced talking signs on vehicles to help people both see and hear the approaching stops, I heard it read out "ST GEORGE ST" as "Street George Street". Sometime later I noticed that they fixed it, but now the lettering read "SNT GEORGE ST". Maybe they figured it was hopeless to teach this computer how to tell the difference between a saint and a street so they just added in a new abbreviation.

  15. David L said,

    September 8, 2022 @ 8:31 pm

    Not long ago I was using Google Maps to navigate around North Wales. She did pretty well with Llandudno (pronouncing the 'u' as a short 'i') but Betwys-y-Coed was beyond her capabilities. Bet-whee-why-ess-why-coed, or something like that.

  16. David L said,

    September 8, 2022 @ 8:33 pm

    I meant Betws-y-coed, Bet-whee-ess-why-coed.

    The friend I was visiting in Wales told me that Welsh is written in a perfectly phonetic fashion, if only you know how to produce the correct sounds.

  17. Richard Sproat said,

    September 8, 2022 @ 11:14 pm

    @David L

    I'll bet the voice didn't pronounce the "Ll" correctly.

  18. David Marjanović said,

    September 9, 2022 @ 3:33 am

    My story about why X gets stressed in X Street is that streets, more so than other types of thoroughfare, are typically found in the vicinity of lots of other streets, and the need to differentiate has led to the institutionalisation of contrastive stress.

    My story about this is that X Street is an ordinary compound noun, stressed on its first part. In earlier centuries spellings such as Oxford-street and Fleet-street were common (at least in London). The stress on the last member in the other kinds of streets, which I didn't know about until now, could then be frozen contrastive stress opposed to "Street".

  19. Ruth Blau said,

    September 9, 2022 @ 8:14 am

    Along a path that I routinely take to get to nearby I-66 is a street named N. Underwood Street. I have to make a left turn there. My GPS says (when I allow her to speak out loud, which is rarely these days), "Turn left on Nor Thunderwood Street." I've just taken my first plunge into owning a smart phone, and I'll be interested to find out how this left turn is pronounced in the various GPS apps available to me.

  20. KeithB said,

    September 9, 2022 @ 8:17 am

    @Kenny
    I wonder why the put the "E" in the street name anyway? Is there another Cesar Chavez that might get confused for Cesar E Chavez?

    Or is it because we always use Martin King Jr's middle name? (See how jarring it is without the "Luther"?

  21. Allan from Iowa said,

    September 9, 2022 @ 8:32 am

    My favorite misreading is East Lincolnway pronounced as East Lean Conway.

  22. Mark P said,

    September 9, 2022 @ 8:33 am

    Do GPS apps know what CYN means? When I used to travel to the LA area I would see exit signs for xxx CYN. It took a while before I realized it was an abbreviation for canyon. I’m sure LA natives know all the canyon roads, so it’s obvious to them, but it wasn’t to me. Kind of like TWP in the northeast.

  23. Hans Adler said,

    September 9, 2022 @ 9:15 am

    The main underlying issue is obviously whether this is (slightly misspelling for clarity) Westwoods Road or West Woodsroad. While the former seems more likely, both are plausible.

    As a human with too much time, I would look on the map for hints. If there is a forest or neighborhood called West Woods, then clearly Westwoods Road is intended. If there is no such thing, but at some intersection West Woods Road turns into East Woods Road (or something similar like North Woods Road), then clearly West Woodsroad is intended. Otherwise I would use a neutral pronunciation consistent with both interpretations.

    The text to speech AI used by Google Maps could conceivably have access to the information required for this decision, but I strongly doubt that it does. I guess that its decisions are based purely on pronunciation patterns in the n-grams in the sentence it is asked to pronounce, compared to those encountered in the training set. Since the latter are probably far from consistent, the chosen pronunciation depends more or less randomly on the surrounding words in each sentence. As a result, it seems to be alternating between West Woodsroad and the neutral pronunciation (or is it Westwoods Road?).

  24. Chris Button said,

    September 9, 2022 @ 9:55 am

    since the English language has irrationally decided to assign primary stress to X in X Street, as opposed to what we do for Road, Avenue, Boulevard, Way, Alley, etc.

    The irrationality is perhaps more on the part of the roads and alleys, etc. than the streets since compounds tend to have single stress.

    John Wells’ Intonation book talks about these confusing “double stressed compounds” and includes a note on how they are treated more regularly in other Germanic languages.

  25. Rodger C said,

    September 9, 2022 @ 2:40 pm

    I've already mentioned somewhere Mountain Pkwy, pronounced Mountain P'kwee.

    And in John Crowley's Little, Big, someone laughs at an early computer program that's printed out "Church of All Streets" and Seventh Saint Bar and Grill," which then later in the novel turn out to be real names.

  26. Coby said,

    September 10, 2022 @ 11:39 am

    My favorite memory of Google Maps navigation is of using my American phone in Quebec City, where 1re (première) Rue was called "one ree roo".

  27. Alyssa Trevelyan said,

    September 10, 2022 @ 7:38 pm

    Interesting! I used to work on this sort of technology – behind all the fancy AI is generally a giant database of street names and their pronunciations, some of which are auto-generated but many of which are hand written. They often have to be done by humans, there's no way to guess from the spelling whether "Rodeo Dr" is going to be ROH-dio or ro-DAY-oh, and there's plenty of both in the US.

    I'd guess that the changing pronunciation is because West Woods Rd ended up divided into several entities in that database. This isn't necessarily a mistake – I used to live near the border of two towns on a street called North Rd, and my home "123 North Rd, Town A" had a counterpart "123 North Rd, Town B" that was only a few blocks down the street. (Very confusing!) It wouldn't be too surprising to see these considered as separate streets that happen to have the same name. In this case, maybe the first stretch of the road ended up with a hand-written pronunciation while the rest has an auto-generated one? Or as earlier suggested, one is generated from "W Woods Rd" and the other from "West Woods Rd".

    The rest of the variation I think is an artifact of sentence-level stress changing, but it's hard to say.

  28. Andy Stow said,

    September 12, 2022 @ 3:18 pm

    It's still a joke in our family after a decade ago using a Garmin to get us to Lakefront Brewery in Milwaukee, WI. As we made the last turn, it announced our destination: lack-ə-front brewery, and for some reason we all just cracked up laughing.

  29. Mac Almy said,

    September 16, 2022 @ 10:03 am

    @Kenny Easwaran

    I had a similar experience of once receiving directions to turn onto "Harry South Truman Parkway."

  30. Rod Johnson said,

    September 20, 2022 @ 7:31 pm

    Mark P: the USPS has a standard set of abbreviations, including CYN. You can find it here. I learned this after I got tangentially involved in a fight about a street sign reading "Parkland Pl," which irate residents pointed out meant "Parkland Place" instead of the correct "Parkland Plaza" ("Plaza" is "Plz").

RSS feed for comments on this post