How well do Google and Microsoft and recognize speech across dialect, gender and race?

If you’ve been following my blog for a while, you may remember that last year I found that YouTube’s automatic captions didn’t work as well for some dialects, or for women. The effects I found were pretty robust, but I wanted to replicate them for a couple of reasons:

  • I only looked at one system, YouTube’s automatic captions, and even that was over a period of several years instead of at just one point in time. I controlled for time-of-upload in my statistical models, but it wasn’t the fairest system evaluation.
  • I didn’t control for the audio quality, and since speech recognition is pretty sensitive to things like background noise and microphone quality, that could have had an effect.
  • The only demographic information I had was where someone was from. Given recent results that find that natural language processing tools don’t work as well for African American English, I was especially interested in looking at automatic speech recognition (ASR) accuracy for African American English speakers.

With that in mind, I did a second analysis on both YouTube’s automatic captions and Bing’s speech API (that’s the same tech that’s inside Microsoft’s Cortana, as far as I know).

Speech Data

For this project, I used speech data from the International Dialects of English Archive. It’s a collection of English speech from all over, originally collected to help actors sound more realistic.

I used speech data from four varieties: the South (speakers from Alabama), the Northern Cities (Michigan), California (California) and General American. “General American” is the sort of news-caster style of speech that a lot of people consider unaccented–even though it’s just as much an accent as any of the others! You can hear a sample here.

For each variety, I did an acoustic analysis to make sure that speakers I’d selected actually did use the variety I thought they should, and they all did.

Systems

For the YouTube captions, I just uploaded the speech files to YouTube as videos and then downloaded the subtitles. (I would have used the API instead, but when I was doing this analysis there was no Python Google Speech API, even though very thorough documentation had already been released.)

Bing’s speech API was a little  more complex. For this one, my co-author built a custom Android application that sent the files to the API & requested a long-form transcript back. For some reason, a lot of our sound files were returned as only partial transcriptions. My theory is that there is a running confidence function for the accuracy of the transcription, and once the overall confidence drops below a certain threshold, you get back whatever was transcribed up to there. I don’t know if that’s the case, though, since I don’t have access to their source code. Whatever the reason, the Bing transcriptions were less accurate overall than the YouTube transcriptions, even when we account for the fact that fewer words were returned.

Results

OK, now to the results. Let’s start with dialect area. As you might be able to tell from the graphs below, there were pretty big differences between the two systems we looked at. In general, there was more variation in the word error rate for Bing and overall the error rate tended to be a bit higher (although that could be due to the incomplete transcriptions we mentioned above). YouTube’s captions were generally more accurate and more consistent. That said, both systems had different error rates across dialects, with the lowest average error rates for General American English.

dialect

Differences in Word Error Rate (WER) by dialect were not robust enough to be significant for Bing (under a one way ANOVA) (F[3, 32] = 1.6, p = 0.21), but they were for YouTube’s automatic captions (F[3, 35] = 3.45,p < 0.05). Both systems had the lowest average WER for General American.

Now, let’s turn to gender. If you read my earlier work, you’ll know that I previously found that YouTube’s automatic captions were more accurate for men and less accurate for women. This time, with carefully recorded speech samples, I found no robust difference in accuracy by gender in either system. Which is great! In addition, the unreliable trends for each system pointed in opposite ways; Bing had a lower WER for male speakers, while YouTube had a lower WER for female speakers.

So why did I find an effect last time? My (untested) hypothesis is that there was a difference in the signal to noise ratio for male and female speakers in the user-uploaded files. Since women are (on average) smaller and thus (on average) slightly quieter when they speak, it’s possible that their speech was more easily masked by background noises, like fans or traffic. These files were all recorded in a quiet place, however, which may help to explain the lack of difference between genders.

gender

Neither Bing (F[1, 34] = 1.13, p = 0.29), nor YouTube’s automatic captions (F[1, 37] = 1.56, p = 0.22) had a significant difference in accuracy by gender.

Finally, what about race? For this part of the analysis, I excluded General American speakers, since they did not report their race. I also excluded the single Native American speaker. Even with fewer speakers, and thus reduced power, the differences between races were still robust enough to be significant for YouTube’s automatic captions and Bing followed the same trend. Both systems were most accurate for Caucasian speakers.

ethnicity

As with dialect, differences in WER between races were not significant for Bing (F[4, 31] = 1.21, p = 0.36), but were significant for YouTube’s automatic captions (F[4, 34] = 2.86,p< 0.05). Both systems were most accurate for Caucasian speakers.

While I was happy to find no difference in performance by gender, the fact that both systems made more errors on non-Caucasian and non-General-American speaking talkers is deeply concerning. Regional varieties of American English and African American English are both consistent and well-documented. There is nothing intrinsic to these varieties that make them less easy to recognize. The fact that they are recognized with more errors is most likely due to bias in the training data. (In fact, Mozilla is currently collecting diverse speech samples for an open corpus of training data–you can help them out yourself.)

So what? Why does word error rate matter?

There are two things I’m really worried about with these types of speech recognition errors. The first is higher error rates seem to overwhelmingly affect already-disadvantaged groups. In the US, strong regional dialects tend to be associated with speakers who aren’t as wealthy, and there is a long and continuing history of racial discrimination in the United States.

Given this, the second thing I’m worried about is the fact that these voice recognition systems are being incorporated into other applications that have a real impact on people’s lives.

Every automatic speech recognition system makes errors. I don’t think that’s going to change (certainly not in my lifetime). But I do think we can get to the point where those error don’t disproportionately affect already-marginalized people. And if we keep using automatic speech recognition into high-stakes situations it’s vital that we get to that point quickly and, in the meantime, stay aware of these biases.

If you’re interested in the long version, you can check out the published paper here.

Advertisements

How many people in the US don’t have an accent?

First, the linguist’s answer: none. Zero. Everyone who uses a language uses a variety of that language, one that reflects their social identity–including things like gender, socioeconomic status or regional background.

But the truth is that some people, especially in the US, have the social privileged of being considered “unaccented”.  I can’t count how many times I’ve been “congratulated” by new acquaintances on having “gotten rid of” my Virginia accent. The thing is, I do have a lot of linguistic features from Tidewater/Piedmont English, like a strong distinction between the vowels in “body” and “baudy”, “y’all” for the second person plural and calling a drive-through liquor store a “brew thru” (shirts with this guy on them were super popular in my high school). But, at the same time, I also don’t have a lot of strongly stigmatized features, like dropping r’s or strong monopthongization you’d hear from a speaker like Virgil Goode (although most folks don’t really sound like that anymore). Plus, I’m young, white, (currently) urban and really highly educated. That, plus the fact that most people don’t pick up on the Southern features I do have, means that I have the privilege of being perceived as accent-less.

You_all_and_Y'all
Map showing the distribution of speakers in the United States who use “y’all”.

But how many people in the US are in the same boat as I am? This is a difficult question, especially given that there is no wide consensus about what “standard”, or “unaccented”, American English is. There is, however, a lot of discussion about what it’s not. In particular, educated speakers from the Midwest and West are generally considered to be standard speakers by non-linguists. Non-linguists also generally don’t consider speakers of African American English and Chicano English to be “standard” speakers (even though both of these are robust, internally consistent language varieties with long histories used by native English speakers).  Fortunately for me, the United States census asks census-takers about their language background, race and ethnicity, educational attainment and geographic location, so I could use census data to roughly estimate how many speakers of “standard” English there are in the United States. I chose to use the 2011 census, as detailed data on language use has been released for that year on a state-by-state basis (you can see a summary here).

From this data, I calculated how many individuals were living in states assigned by the U.S. Census Bureau to either the West or Midwest and how many residents surveyed in these states reported speaking English ‘very well’ or better. Then, assuming that residents of these states had educational attainment rates representative of national averages, I estimated how many college educated (with a bachelor’s degree or above) non-Black and non-Hispanic speakers lived in these areas.

So just how many speakers fit into this “standard” mold? Fewer than you might expect! You can see the breakdown below:

Speakers in the 2011 census who…

Count

% of US Population

…live in the United States…

311.7 million

100%

…and live in the Midwest or West…

139,968,791

44.9%

…and speak English at least ‘very well’…

127,937,178

41%

…and are college educated…

38,381,153 (estimated)

12.31%

…and are not Black or Hispanic.

33,391,603 (estimated)

10.7%

Based on the criteria laid out above, only around a tenth of the US population would count as ‘standard’ speakers. Now, keep in mind this estimate is possibly somewhat conservative: not all Black speakers use African American English and not all Hispanic speakers use Chicano English, and the regional dialects of some parts of the Northeast are also sometimes considered “standard”, which isn’t reflected in my rough calculation. That said, I think there’s still something if a large majority of Americans don’t speak what we might consider “standard” English, maybe it’s time to start redefining who gets to be the standard.

Can what you think you know about someone affect how you hear them?

I’ll get back to “a male/a female” question in my next blog post (promise!), but for now I want to discuss some of the findings from my dissertation research. I’ve talked about my dissertation research a couple times before, but since I’m going to be presenting some of it in Spain (you can read the full paper here), I thought it would be a good time to share some of my findings.

In my dissertation, I’m looking at how what you think you know about a speaker affects what you hear them say. In particular, I’m looking at American English speakers who have just learned to correctly identify the vowels of New Zealand English. Due to an on-going vowel shift, the New Zealand English vowels are really confusing for an American English speaker, especially the vowels in the words “head”, “head” and “had”.

tokensVowelPlot

This plot shows individual vowel tokens by the frequency of thier first and second formants (high-intensity frequency bands in the vowel). Note that the New Zealand “had” is very close to the US “head”, and the New Zealand “head” is really close to the US “hid”.

These overlaps can be pretty confusing when American English speakers are talking to New Zealand English speakers, as this Flight of the Conchords clip shows!

The good news is that, as language users, we’re really good at learning new varieties of languages we already know, so it only takes a couple minutes for an American English speaker to learn to correctly identify New Zealand English vowels. My question was this: once an American English speaker has learned to understand the vowels of New Zealand English, how do they know when to use this new understanding?

In order to test this, I taught twenty one American English speakers who hadn’t had much, if any, previous exposure to New Zealand English to correctly identify the vowels in the words “head”, “heed” and “had”. While I didn’t play them any examples of a New Zealand “hid”–the vowel in “hid” is said more quickly in addition to having different formants, so there’s more than one way it varies–I did let them say that they’d heard “hid”, which meant I could tell if they were making the kind of mistakes you’d expect given the overlap between a New Zealand “head” and American “hid”.

So far, so good: everyone quickly learned the New Zealand English vowels. To make sure that it wasn’t that they were learning to understand the one talker they’d been listening to, I tested half of my listeners on both American English and New Zealand English vowels spoken by a second, different talker. These folks I told where the talker they were listening to was from. And, sure enough, they transferred what they’d learned about New Zealand English to the new New Zealand speaker, while still correctly identifying vowels in American English.

The really interesting results here, though, are the ones that came from the second half the listeners. This group I lied to. I know, I know, it wasn’t the nicest thing to do, but it was in the name of science and I did have the approval of my institutional review board, (the group of people responsible for making sure we scientists aren’t doing anything unethical).

In an earlier experiment, I’d played only New Zealand English as this point, and when I told them the person they were listening to was from America, they’d completely changed the way they listened to those vowels: they labelled New Zealand English vowels as if they were from American English, even though they’d just learned the New Zealand English vowels. And that’s what I found this time, too. Listeners learned the New Zealand English vowels, but “undid” that learning if they thought the speaker was from the same dialect as them.

But what about when I played someone vowels from their own dialect, but told them the speaker was from somewhere else? In this situation, listeners ignored my lies. They didn’t apply the learning they’d just done. Instead, the correctly treated the vowels of thier own dialect as if they were, in fact, from thier dialect.

At first glance, this seems like something of a contradiction: I just said that listeners rely on social information about the person who’s talking, but at the same time they ignore that same social information.

So what’s going on?

I think there are two things underlying this difference. The first is the fact that vowels move. And the second is the fact that you’ve heard a heck of a lot more of your own dialect than one you’ve been listening to for fifteen minutes in a really weird training experiment.

So what do I mean when I say vowels move? Well, remember when I talked about formants above? These are areas of high acoustic energy that occur at certain frequency ranges within a vowel and they’re super important to human speech perception. But what doesn’t show up in the plot up there is that these aren’t just static across the course of the vowel–they move. You might have heard of “diphthongs” before: those are vowels where there’s a lot of formant movement over the course of the vowel.

And the way that vowels move is different between different dialects. You can see the differences in the way New Zealand and American English vowels move in the figure below. Sure, the formants are in different places—but even if you slid them around so that they overlapped, the shape of the movement would still be different.

formantDynamics

Comparison of how the New Zealand and American English vowels move. You can see that the shape of the movement for each vowel is really different between these two dialects.  

Ok, so the vowels are moving in different ways. But why are listeners doing different things between the two dialects?

Well, remember how I said earlier that you’ve heard a lot more of your own dialect than one you’ve been trained on for maybe five minutes? My hypothesis is that, for the vowels in your own dialect, you’re highly attuned to these movements. And when a scientist (me) comes along and tells you something that goes against your huge amount of experience with these shapes, even if you do believe them, you’re so used to automatically understanding these vowels that you can’t help but correctly identify them. BUT if you’ve only heard a little bit of a new dialect you don’t have a strong idea of what these vowels should sound like, so if you’re going to rely more on the other types of information available to you–like where you’re told the speaker is from–even if that information is incorrect.

So, to answer the question I posed in the title, can what you think you know about someone affect how you hear them? Yes… but only if you’re a little uncertain about what you heard in the first place, perhaps becuase it’s a dialect you’re unfamiliar with.

Tweeting with an accent

I’m writing this blog post from a cute little tea shop in Victoria, BC. I’m up here to present at the Northwest Linguistics Conference, which is a yearly conference for both Canadian and American linguists (yes, I know Canadians are Americans too, but United Statsian sounds weird), and I thought that my research project may be interesting to non-linguists as well. Basically, I investigated whether it’s possible for Twitter users to “type with an accent”. Can linguists use variant spellings in Twitter data to look at the same sort of sound patterns we see in different speech communities?

Picture of a bird saying

Picture of a bird saying “Let’s Tawk”. Taken from the website of the Center for the Psychology of Women in Seattle. Click for link.

So if you’ve been following the Great Ideas in Linguistics series, you’ll remember that I wrote about sociolinguistic variables a while ago. If you didn’t, sociolinguistic variables are sounds, words or grammatical structures that are used by specific social groups. So, for example, in Southern American English (representing!) the sound in “I” is produced with only one sound, so it’s more like “ah”.

Now, in speech these sociolinguistic variables are very well studied. In fact, the Dictionary of American Regional English was just finished in 2013 after over fifty years of work. But in computer mediated communication–which is the fancy term for internet language–they haven’t been really well studied. In fact, some scholars suggested that it might not be possible to study speech sounds using written data. And on the surface of it, that does make sense. Why would you expect to be able to get information about speech sounds from a written medium? I mean, look at my attempt to explain an accent feature in the last paragraph. It would be far easier to get my point across using a sound file. That said, I’d noticed in my own internet usage that people were using variant spellings, like “tawk” for “talk”, and I had a hunch that they were using variant spellings in the same way they use different dialect sounds in speech.

While hunches have their place in science, they do need to be verified empirically before they can be taken seriously. And so before I submitted my abstract, let alone gave my talk, I needed to see if I was right. Were Twitter users using variant spellings in the same way that speakers use different sound patterns? And if they are, does that mean that we can investigate sound  patterns using Twitter data?

Since I’m going to present my findings at a conference and am writing this blog post, you can probably deduce that I was right, and that this is indeed the case. How did I show this? Well, first I picked a really well-studied sociolinguistic variable called the low back merger. If you don’t have the merger (most African American speakers and speakers in the South don’t) then you’ll hear a strong difference between the words “cot” and “caught” or “god” and “gaud”. Or, to use the example above, you might have a difference between the words “talk” and “tock”. “Talk” is little more backed and rounded, so it sounds a little more like “tawk”, which is why it’s sometimes spelled that way. I used the Twitter public API and found a bunch of tweets that used the “aw” spelling of common words and then looked to see if there were other variant spellings in those tweets. And there were. Furthermore, the other variant spellings used in tweets also showed features of Southern American English or African American English. Just to make sure, I then looked to see if people were doing the same thing with variant spellings of sociolinguistic variables associated with Scottish English, and they were. (If you’re interested in the nitty-gritty details, my slides are here.)

Ok, so people will sometimes spell things differently on Twitter based on their spoken language dialect. What’s the big deal? Well, for linguists this is pretty exciting. There’s a lot of language data available on Twitter and my research suggests that we can use it to look at variation in sound patterns. If you’re a researcher looking at sound patterns, that’s pretty sweet: you can stay home in your jammies and use Twitter data to verify findings from your field work. But what if you’re not a language researcher? Well, if we can identify someone’s dialect features from their Tweets then we can also use those features to make a pretty good guess about their demographic information, which isn’t always available (another problem for sociolinguists working with internet data). And if, say, you’re trying to sell someone hunting rifles, then it’s pretty helpful to know that they live in a place where they aren’t illegal. It’s early days yet, and I’m nowhere near that stage, but it’s pretty exciting to think that it could happen at some point down the line.

So the big take away is that, yes, people can tweet with an accent, and yes, linguists can use Twitter data to investigate speech sounds. Not all of them–a lot of people aren’t aware of many of their dialect features and thus won’t spell them any differently–but it’s certainly an interesting area for further research.