Beyond the Every Day: Vocal Potential in AI Mediated Communication
In summer 2021, sound artist, engineer, musician, and educator Johann Diedrick convened a panel at the intersection of racial bias, listening, and AI technology at Pioneerworks in Brooklyn, NY. Diedrick, 2021 Mozilla Creative Media award recipient and creator of such works as Dark Matters, is currently working on identifying the origins of racial bias in voice interface systems. Dark Matters, according to Squeaky Wheel, “exposes the absence of Black speech in the datasets used to train voice interface systems in consumer artificial intelligence products such as Alexa and Siri. Utilizing 3D modeling, sound, and storytelling, the project challenges our communities to grapple with racism and inequity through speech and the spoken word, and how AI systems underserve Black communities.” And now, he’s working with SO! as guest editor for this series for Sounding Out! (along with ed-in-chief JS!). It starts today, with Amina Abbas-Nazari, helping us to understand how Speech AI systems operate from a very limiting set of assumptions about the human voice– are we training it, or is it actually training us?
—
Hi, good morning. I’m calling in from Bangalore, India.” I’m talking on speakerphone to a man with an obvious Indian accent. He pauses. “Now I have enabled the accent translation,” he says. It’s the same person, but he sounds completely different: loud and slightly nasal, impossible to distinguish from the accents of my friends in Brooklyn.
The AI startup erasing call center worker accents: is it fighting bias – or perpetuating it? (Wilfred Chan, 24 August 2022)
This telephone interaction was recounted in The Guardian reporting on a Silicon Valley tech start-up called Sanas. The company provides AI enabled technology for real-time voice modification for call centre workers voices to sound more “Western”. The company describes this venture as a solution to improve communication between typically American callers and call centre workers, who might be based in countries such as Philippines and India. Meanwhile, research has found that major companies’ AI interactive speech systems exhibit considerable racial imbalance when trying to recognise Black voices compared to white speakers. As a result, in the hopes of being better heard and understood, Google smart speaker users with regional or ethnic American accents relay that they find themselves contorting their mouths to imitate Midwestern American accents.
These instances describe racial biases present in voice interactions with AI enabled and mediated communication systems, whereby sounding ‘Western’ entitles one to more efficient communication, better usability, or increased access to services. This is not a problem specific to AI though. Linguistics researcher John Baugh, writing in 2002, describes how linguistic profiling is known to have resulted in housing being denied to people of colour in the US via telephone interactions. Jennifer Stoever‘s The Sonic Color Line (2016) presents a cultural and political history of the racialized body and how it both informed and was informed by emergent sound technologies. AI mediated communication repeats and reinforces biases that pre-exist the technology itself, but also helping it become even more widely pervasive.
Mozilla’s commendable Common Voice project aims to ‘teach machines how real people speak’ by building an open source, multi-language dataset of voices to improve usability for non-Western speaking or sounding voices. But singer and musicologist, Nina Sun Eidsheim describes how ’a specific voice’s sonic potentiality [in] its execution can exceed imagination’ (7), and voices as having ‘an infinity of unrealised manifestations’ (8) in The Race of Sound (2019). Eidsheim’s sentiments describe a vocal potential, through musicality, that exists beyond ideas of accents and dialects, and vocal markers of categorised identity. As a practicing vocal performer, I recognise and resonate with Eidsheim’s ideas I have a particular interest in extended and experimental vocality, especially gained through my time singing with Musarc Choir and working with artist Fani Parali. In these instances, I have experienced the pleasurable challenge of being asked to vocalise the mythical, animal, imagined, alien and otherworldly edges of the sonic sphere, to explore complex relations between bodies, ecologies, space and time, illuminated through vocal expression.
Following from Eidsheim, and through my own vocal practice, I believe AI’s prerequisite of voices as “fixed, extractable, and measurable ‘sound object[s]’ located within the body” is over-simplistic and reductive. Voices, within systems of AI, are made to seem only as computable delineations of person, personality and identity, constrained to standardised stereotypes. By highlighting vocal potential, I offer a unique critique of the way voices are currently comprehended in AI recognition systems. When we appreciate the voice beyond the homogenous, we give it authority and autonomy, ultimately leading to a fuller understanding of the voice and its sounding capabilities.
My current PhD research, Speculative Voicing, applies thinking about the voice from a musical perspective to the sound and sounding of voices in artificially intelligent conversational systems. Herby the voice becomes an instrument of the body to explore its sonic materiality, vocal potential and extremities of expression, rather than being comprehended in conjunction to vocal markers of identity aligning to categories of race, gender, age, etc. In turn, this opens space for the voice to be understood as a shapeshifting, morphing and malleable entity, with immense sounding potential beyond what might be considered ordinary or everyday speech. Over the long term this provides discussion of how experimenting with vocal potential may illuminate more diverse perspectives about our sense of self and being in relation to vocal sounding.
Vocal and movement artist Elaine Mitchener exhibits the disillusion of the voice as ‘fixed’ perfectly in her performance of Christian Marclay’s No!, which I attended one hot summer’s evening at the London Contemporary Music Festival in 2022. Marclay’s graphic score uses cut outs from comic book strips to direct the performer to vocalise a myriad of ‘No”s.
Mitchener’s rendering of the piece involved the cooperation and coordination of her entire body, carefully crafting lips, teeth, tongue, muscles and ligaments to construct each iteration of ‘No.’ Each transmutation of Mitchener’s ‘No’s’ came with a distinct meaning, context, and significance, contained within the vocalisation of this one simple syllable. Every utterance explored a new vocal potential, enabled by her body alone. In the context of AI mediated communication, we can see this way of working with the voice renders the idea of the voice as ‘fixed’ as redundant. Mitchener’s vocal potential demonstrates that voices can and do exist beyond AI’s prescribed comprehension of vocal sounding.
In order to further understand how AI transcribes understandings of voice onto notions of identity, and vocal potential, I produced the practice project Polyphonic Embodiment(s) as part of my PhD research, in collaboration with Nestor Pestana, with AI development by Sitraka Rakotoniaina. The AI we created for this project is based upon a speech-to-face recognition AI that aims to be able to tell what your face looks like from the sound of your voice. The prospective impact of this AI is deeply unsettling, as its intended applications are wide-ranging – from entertainment to security, and as previously described AI recognition systems are inherently biased.
This multi-modal form of comprehending voice is also a hot topic of research being conducted by major research institutions including Oxford University and Massachusetts Institute of Technology. We wanted to explore this AI recognition programme in conjunction with an understanding of vocal potential and the voice as a sonic material shaped by the body. As the project title suggests, the work invites people to consider the multi-dimensional nature of voice and vocal identity from an embodied standpoint. Additionally, it calls for contemplation of the relationships between voice and identity, and individuals having multiple or evolving versions of identity. The collaboration with the custom-made AI software creates a feedback loop to reflect on how peoples’ vocal sounding is “seen” by AI, to contest the way voices are currently heard, comprehended and utilised by AI, and indeed the AI industry.
The video documentation for this project shows ‘facial’ images produced by the voice-to-face recognition AI, when activated by my voice, modified with simple DIY voice devices. Each new voice variation, created by each device, produces a different outputted face image. Some images perhaps resemble my face? (e.g. Device #8) some might be considered more masculine? (e.g. Device #10) and some are just disconcerting (e.g. Device #4). The speculative nature of Polyphonic Embodiment(s) is not to suggest that people should modify their voices in interaction with AI communication systems. Rather the simple devices work with bodily architecture and exaggerate its materiality, considering it as a flexible instrument to explore vocal potential. In turn this sheds light on the normative assumptions contained within AI’s readings of voice and its relationships to facial image and identity construction.
Through this artistic, practice-led research I hope to evolve and augment discussion around how the sounding of voices is comprehended by different disciplines of research. Taking a standpoint from music and design practice, I believe this can contest ways of working in the realms of AI mediated communication and shape the ways we understand notions of (vocal) identity: as complex, fluid, malleable, and ultimately not reducible to Western logics of sounding.
—
Featured Image: Still image from Polyphonic Embodiments, courtesy of author.
—
Amina Abbas-Nazari is a practicing speculative designer, researcher, and vocal performer. Amina has researched the voice in conjunction with emerging technology, through practice, since 2008 and is now completing a PhD in the School of Communication at the Royal College of Art, focusing on the sound and sounding of voices in artificially intelligent conversational systems. She has presented her work at the London Design Festival, Design Museum, Barbican Centre, V&A, Milan Furniture Fair, Venice Architecture Biennial, Critical Media Lab, Switzerland, Litost Gallery, Prague and Harvard University, America. She has performed internationally with choirs and regularly collaborates with artists as an experimental vocalist
—
REWIND! . . .If you liked this post, you may also dig:
What is a Voice?–Alexis Deighton MacIntyre
Mr. and Mrs. Talking Machine: The Euphonia, the Phonograph, and the Gendering of Nineteenth Century Mechanical Speech – J. Martin Vest
One Scream is All it Takes: Voice Activated Personal Safety, Audio Surveillance, and Gender Violence—María Edurne Zuazu
Echo and the Chorus of Female Machines—AO Roberts
On Sound and Pleasure: Meditations on the Human Voice– Yvon Bonefant
Share this:
Voice as Ecology: Voice Donation, Materiality, Identity
I first heard about voice donation while listening to “Being Siri,” an experimental audio piece about Erin Anderson donating her voice to Boston-based voice donation company, VocaliD. Like a digital blood bank of sorts, VocaliD provides a platform for donating one’s voice via digital audio recordings. These recordings are used to help technicians create a custom digital voice for a voiceless individual, providing an alternative to the predominately white, male, mechanical-sounding assistive technologies used by people who cannot vocalize for themselves (think Stephen Hawking). VocaliD manufactures voices that better match a person’s race, gender, ethnicity, age, and unique personality. To me, VocaliD encapsulates the promise, complexity, and problematic nature of our current speech AI landscape and serves as an example of why we need to think critically about sound technologies, even when they appear to be wholly beneficial.
Given the extreme lack of sonic diversity in vocal assistive technologies, VocaliD provides a critically important service. But a closer look at both the rhetoric used by the organization and the material process involved in voice donation also amplifies the limits of overly simplistic, human-centric conceptions of voice. For instance, VocaliD rhetorically frames their service by persistently linking voice to humanity—to self, authenticity, individuality. Consider the following statements made by Rupal Patel, CEO and founder of VocaliD, in which she emphasizes the need for voice donation technology:
“Here’s a way for us to acknowledge these individuals as unique human beings.” (Fast Company)
“I was talking to [a] girl we made a voice for. She told me that people are finally seeing her for who she really is.” (Medieros)
These are just a few examples from a larger discourse that reinforces the connection between voice and humanity. VocaliD’s repeated claims that their unique vocal identities humanize individuals imply that one is not fully human unless one’s voice sounds human. This rhetoric positions voiceless individuals as less than human (at least until they pay for a customized human-sounding voice).
VocaliD’s conflation of voice and humanity makes me wonder about the meaning of “human” in this context. For example, notions of humanity have been historically associated with Western whiteness—and deployed as a means of separating or distinguishing white people from Others—as Alexander Weheliye points out. Though VocaliD’s mission is to diversify manufactured voices, is a “human-sounding” voice still construed as a white voice? Does sounding human mean sounding white? Even if there is a bank of sonically diverse voices to choose from, does racial bias show up in the pacing, phrasing, or inflection caused by the vocal technology?
I am also disturbed by the rhetoric of humanity and individuality used by VocaliD because the company adopts the same rhetoric to describe the AI voices they sell to brands for media and smart products. Here’s an example of this rhetoric from the VocaliD AI website: “When you need a voice that resonates, evokes audience empathy, and sounds like you, rather than your competitors, VocaliD’s AI-powered vocal persona is the solution. Your voice — always on, where you need it when you need it.” Using similar rhetorical strategies to describe both voiceless people and products is dehumanizing. And yet, having a more diverse AI vocal mediascape, especially in terms of race, is crucially important since voice-activated machines and products are designed largely by white men who end up reinforcing the sonic color line.
Interestingly, the processes VocaliD uses to create a custom voice reveal that these voices are not, in fact, unique markers of humanity or individuality. It’s hard to find a detailed account of how VocaliD voices are made due to the company’s patents, but here are the basics: VocaliD does not transfer a donated voice directly to a voiceless person’s assistive technology. VocaliD technicians instead blend and digitally manipulate the donated voice with recordings of the noises a voiceless person can make (a laugh, a hum) to create a distinct new voice for the recipient. In other words, donated voices are skillful remixes that wouldn’t be possible without extracting vocal data and manipulating it with digital tools. Despite perpetuating narratives about voice, humanity, and authenticity, VocaliD’s creative blending of vocal material reveals that donated voices are the result of compositional processes that involve much more than people.
Further, considering VocaliD voices from a material rather than human-centric perspective amplifies something important about voices in general. All voices are composed of and grounded in an ecology. That is, voices emerge and are developed through a mixture of: (1) biological makeup (or technological makeup in the case of machines with voices); (2) specific environments and contexts (geography may determine the kind of accents humans have; AI voices have distinct sounds for their brands); (3) technologies (phones, computers, digital recorders and editors, software, and assistive technologies preserve, circulate, and amplify voices); and (4) others (humans often emulate the vocal patterns of the people they interact with most; many machine voices also sound like other machine voices). Put simply, all voices are intentionally and unintentionally composed over time—shaped by ever-changing bodily (and/or technological) states and engagements with the world. Voices are dynamic compositions by nature. Examining voice from a material standpoint shows that voices are not static markers of humanity; voices are responsive and malleable because they are the result of a complex ecology that involves much more than a “unique” human being.
However, focusing solely on the material aspects of vocality leaves out people’s lived experiences of voice. And based on online videos of VocaliD recipients—like Delaney, a seventeen-year-old with cerebral palsy—VocaliD voices seem to live up to the company’s hype. Delaney appears delighted by her new voice, stating: “I was so excited to get my own voice. I used to have a computer voice and now I sound like a girl. I like that. And I talk more.” Delaney’s teachers also discuss how her new voice completely changed her demeanor. Whereas before Delaney was reluctant to use her assistive technology to speak, her new voice gives her confidence and a stronger sense of identity. As her teacher explains in the video, “she is really engaged in groups, she wants to share her answers, she’s excited to talk with friends. It’s been really nice to see.” For Delaney, a VocaliD voice represents a newfound sense of agency.
It’s important to recognize this video is not necessarily representative of every VocaliD recipient’s experience, or even Delaney’s full experience. As Meryl Alper notes in Giving Voice, these types of news stories “portray technology as allowing individuals to ‘overcome’ their disability as an individual limitation, and are intended to be uplifting and inspirational for able-bodied audiences” (27). While we should be wary of the technological determinism in the video, observing Delaney use her VocaliD voice—and listening to the emotional responses of her mom and teachers—makes it difficult to deny that donated voices make a positive impact. For me, this video also gets at a larger truth about humans and voice: the ways we hear and understand our own voices, and the ways others interpret the sounds of our voices, matter a great deal. Voices are integral to our identities—to the ways we understand and think about ourselves and others—and the sounds of our voices have social and material consequences, as the SO! Gendered Voices Forum illustrates so clearly.
It’s worth repeating that VocaliD’s mission to diversify synthetic voices is incredibly important, especially given the restrictive vocal options available to voiceless individuals. It’s also necessary to acknowledge the company has limitations that end up reproducing the structural inequities it tries to address. As Alper observes, “In order to become a speech donor, one must have three to four hours of spare time to record their speech, access to a steady and strong Internet connection, and a quiet location in which to record” (162-63). With these obstacles to donating one’s voice in mind, it’s not surprising that all the VocaliD recipient videos I could find feature white people. Donating one’s voice is much easier for middle to upper class white people who have access to privacy, Internet, and leisure time.
This brief examination of VocaliD raises questions about what a more equitable future for vocal technologies might look/sound like. Though I don’t have the answer, I believe that to understand the fullness of voice, we can’t look at it from a single perspective. We need to account for the entire vocal ecology: the material (biological, technological, financial, etc.) conditions from which a voice emerges or is performed, and individual speakers’ understanding of their culture, race, ethnicity, gender, class, ability, sexuality, etc. An ecological approach to voice involves collaborating with people and their vocal needs and desires—something VocaliD models already. But it also involves accounting for material realities: How might we make the barriers preventing a more diverse voice ecosystem less difficult to navigate—especially for underrepresented groups? In short, we must treat voice holistically. Voices are more than people, more than technologies, more than contexts, more than sounds. Understanding voice means acknowledging the interconnectedness of these things and how that interconnectedness enables or precludes vocal possibilities.
—
Featured image: 366-350 You can’t shut me up, Jennifer Moo, CC BY-ND
—
Steph Ceraso is an associate professor of digital writing and rhetoric at the University of Virginia. Her 2018 book, Sounding Composition: Multimodal Pedagogies for Embodied Listening, proposes an expansive approach to teaching with sound in the composition classroom. She also published a digital book in 2019 called Sound Never Tasted So Good: ‘Teaching’ Sensory Rhetorics—an exploration of writing, sound, rhetoric, and food. She is currently working on a book project that examines sonic forms of invention in various contexts.
—
REWIND! . . .If you liked this post, you may also dig:
What is a Voice?–Alexis Deighton MacIntyre
Mr. and Mrs. Talking Machine: The Euphonia, the Phonograph, and the Gendering of Nineteenth Century Mechanical Speech – J. Martin Vest
Only the Sound Itself?: Early Radio, Education, and Archives of “No-Sound”–Amanda Keeler
Share this:
Current Top Posts
April 2021
January 2020
November 2019
October 2019
August 2019
April 2014
November 2017
March 2015
December 2017
October 2015
Sounding Out! 
Recent Comments