This text was originally published as a topic essay in the Curriculum. An archival version with embedded audio is available here, and in German translation here. But we have also extracted the text for the purposes of legibility. [formatting is a work in progress]
“Machine listening” is one common term for a fast-growing interdisciplinary field of science and engineering which uses audio signal processing and machine learning to “make sense” of sound and speech.1Machine listening is what enables you to be “understood” by Siri and Alexa, to Shazam a song, and to interact with many audio-assistive technologies if you are blind or vision impaired.2 As early as the 90s, the term was already being used in computer music to describe the analytic dimension of www⁄‘interactive music systems’3 whose behavior changes in response to live musical input, though there are precedents even before that.4 Machine Listening was also, of course, a cornerstone of the mass surveillance programs revealed by Edward Snowden in 2013: SPIRITFIRE’s “speech-to-text keyword search and paired dialogue transcription”; EViTAP’s “automated news monitoring”; VoiceRT’s “ingestion”, according to one NSA slide, of Iraqi voice data into voiceprints. Domestically, machine listening technologies underpin the vast databases of vocal biometrics now held by many www⁄prison providers and, for instance, the www⁄Australian Tax Office. And they are quickly being integrated into infrastructures of development, security and policing.
Automatic speech recognition,5 transcription and translation [i, ii, iii] - targeted key word detection [i] - vocal biometrics and audio fingerprinting6[i, ii]- speaker identification, differentiation, enumeration and location [i, ii] - personality and emotion recognition [i] - accent identification [i] - sound recognition [i]- audio object recognition [i] - auditory scene analysis [i] - intelligent audio analysis7 - audio event analysis8 - audio context awareness [i] - music mood analysis [i] - music identification [i] - cover song identification [i] - music playlist generation9 - audio synthesis [i] - speech synthesis [i, ii, iii] - musical synthesis [i] - adversarial music [i] - brand sonification [i] - aggression detection [i, ii] - depression detection [i] - laughter detection [i] - emotion detection - [i] intoxication detection [i] - scream detection10 - lie detection [i] - hoax detection [i] - gunshot detection - [i] parkinson’s diagnosis [i] - covid diagnosis [i] - machine fault diagnosis - psychosis diagnosis [i] - bird sound identification [i] - gender identification and ethnicity detection [i] - age determination [i] - voice likability determination [i] - risk assessment [i]…
These applications are all either currently in use by states, corporations and other entities around the world, or under development. The list is obviously not exhaustive.11 Nor does it convey the real diversity of markets, cyberphysical, social and political contexts into which these applications are quickly embedding themselves:
Digital voice assistants - voice user interfaces - state and corporate surveillance [i] - profiling - border security - home security - pre-emptive policing - weapons systems - court systems [i, ii] - hospital systems - call centre optimisation - oral hygiene [i] - disability services - grocery store wayfinding [i] - ambient elderly monitoring [i] - baby monitoring [i]- house arrest monitoring [i] - human rights monitoring12 - remote education and proctoring [i]- school security [i] - remote diagnostics - biomonitoring and personalised health [i] - social distancing - music streaming - music education - composition [i] - gaming - [i] brand development [i] - marketing [i] - acoustic ecology [i] - employee performance metrics [i] - wearables [i] - hearables [i, ii] - recruitment, banking and insurance [i] - voice banking and vocal proshetics [i]
As with all forms of machine learning, questions of efficacy, access, privacy, bias, fairness and transparency arise with every use case. But machine listening also demands to be treated as an epistemic and political system in its own right, that www⁄casts a shadow, that increasingly enables, shapes and constrains basic human possibilities, that is making our auditory worlds knowable in new ways, to new institutions, according to new logics, and is remaking (sonic) life in the process.
Machine listening is much more than just a new scientific discipline or vein of technical innovation then. It is also an emergent field of knowledge-power and cultural production, of data extraction and colonialism, of capital accumulation, automation and control. We must make it a field of political contestation and struggle. If there is to be a world of listening machines, we must make it emancipatory.13
Machine listening¶
Machine listening isn’t just machinic.
Materially, it entails enormous exploitation of both human and planetary resources: to build, power and maintain the vast infrastructures on which it depends, along with all the microphones and algorithms which are its www⁄most visible manifestations.14 Even these are not so visible however. One of the many political challenges machine listening presents is its tendency to disappear at point of use, even as it indelibly marks the bodies of distant workers and permanently scars ecological systems.
Scientifically, machine listening demands enormous volumes of data: exhorted, extracted and appropriated from auditory environments and cultures which, though numerous already, will never be diverse enough. This is why responding to machinic bias with a politics of inclusion can also be a trap.15 It means committing to the very system that is oppressing or occluding you: a “techno-politics of perfection."16
Because machine listening is trained on (more-than) human auditory worlds, it inevitably encodes, invisibilises and reinscribes normative listenings, along with a range of more arbitrary artifacts of the datasets, statistical models and computational systems which are at once its lifeblood and fundamentally opaque.17 This combination means that machine listening is simultaneously an alibi or front for the proliferation and normalisation of specific auditory practices as machinic, and, conversely, often irreducible to human apprehension; which is to say the worst of both worlds.
Moreover, because machine listening is so deeply bound up with logics of automation and pre-emption, it is also recursive. It feeds its listenings back into the world - gendered and gendering,18 colonial and colonizing, raced and racializing,15 classed and productive of class relations - as Siri’s answer or failure to answer; by alerting the police, denying your www⁄claim for asylum, or continuing to play Autechre - and this incites an auditory response to which it listens in turn. The soundscape is increasingly cybernetic. Confronting machine listening means recognising that common-sense distinctions between human and machine simply fail to hold. We are all machine listeners now. We have been becoming machine listeners for a long time. Indeed, the becoming machinic of listening is a foundational concern for any contemporary politics of listening; not because mechanisation itself is a problem, but because it is the condition in which we increasingly find ourselves.19
But machine listening isn’t exactly listening either.
Technically, the methods of machine listening are diverse, but they bear little relationship to the biological processes of human audition or psychocultural processes of meaning making. Many are fundamentally www⁄imagistic, in the sense that they work by first transforming sound into spectograms. Many work by combining auditory with other forms of data and sensory inputs: machines that www⁄listen by looking, or by cross-referencing audio with geolocation data. In the field of Automatic Speech Recognition, for instance, it was only when researchers at IBM moved away from attempts to simulate human listening towards statistical data processing in the 1970s that the field began making decisive steps forward.20 Speech recognition needed to untether itself from “human sensory-motor phenomenon” in order to start recognising speech. Airplanes don’t flap their wings. 20
Even if machine listening did work by analogising human audition, the question of cognition would still remain. Insofar as “listening” implies a subjectivity, machines do not (yet) listen. But this kind of anthropocentrism simply begs the question. What is at stake with machine listening is precisely a new auditory regime: an analogue of Paul Virilio’s “sightless vision”, 21 the possibility of a listening without hearing or comprehension, a purely correlative listening, with the human subject decentered as privileged auditor.
One way of responding to this possibility would be to simply bracket the question of listening and think in terms of “listening effects” instead, so that the question is no longer whether machines are listening, but what it means to live in a world in which they act like it, and we do too.
Another response would be to say that when or if machines listen, they listen operationally: not in order to understand, or even facilitate human understanding, but to perform an operation: to diagnose, to identify, to recognize, to trigger.22 And we could notice that as listening becomes increasingly operational sound does too. www⁄Operational acoustics: sounds made by machines for machine listeners. www⁄Adversarial acoustics: sounds made by machines against human listeners, and vice versa.23
Resources¶
- ⦚bib:7c769ce6-5e9e-40d3-96ef-1838a7f57365not found ↩︎
- Meryl Alper, Giving Voice: Mobile Communication, Disability, and Inequality (MIT Press, 2017) ↩︎
- Robert Rowe, www⁄Interactive music systems: Machine listening and composing.Cambridge, MA: The MIT Press (1993) ↩︎
- Stefan Maier, www⁄Machine Listening, Technosphere Magazine (2018); Interview with www⁄Stefan Maier on September 11, 2020 ↩︎
- Interview with www⁄Kathy Reid on August 11, 2020 ↩︎
- Xiaochang Li and Mara Mills, “Vocal Features: From Identification to Speech Recognition by Machine” 60(2) Technology and Culture (2019) pp.129-S160 DOI: https://doi.org/10.1353/tech.2019.0066 ↩︎
- ⦚bib:827d1f44-5a35-4278-a527-4df67e5ba321not found ↩︎
- ⦚bib:7cf99c5d-1a28-44d9-958a-8ff5e9cb4441not found ↩︎
- ⦚bib:d2b1e24c-c800-42b9-ba67-105b0b25efc9not found ↩︎
- Lei, B., Mak, MW. “Robust scream sound detection via sound event partitioning. Multimed Tools Appl” 75, 6071–6089 (2016). https://doi.org/10.1007/s11042-015-2555-z ↩︎
- Interview with www⁄Shannon Mattern on August 18, 2020 ↩︎
- Interview with www⁄André Dao on September 4, 2020 ↩︎
- See for instance www⁄Data 4 Black Lives, www⁄Feminist Data Manifest-No ↩︎
- ⦚bib:3f8dd486-3e28-45ef-929f-65086850870enot found ↩︎
- Interview with www⁄Halcyon Lawrence on August 31, 2020. See also Thao Phan, “Amazon Echo and the Aesthetics of Whiteneness” 5(1) Catalyst: Feminism, Theory, Technoscience (2019), 1-38. ↩︎
- ⦚bib:6e8f7c36-d251-4a07-ac5d-0b938c5f5feenot found ↩︎
- ⦚bib:c58be9a5-a599-4a4b-b58f-a07721fc1721not found ↩︎
- ⦚bib:26f7b730-9064-464b-b905-fbe63c5d4e4bnot found ↩︎
- Interview with Lawrence Abu Hamdan, publication forthcoming 2021 ↩︎
- ⦚bib:6676af8a-7a4d-4aa8-af96-f26452f58753not found ↩︎
- ⦚bib:8558647f-101d-43ff-a531-5df8eb87199anot found p.53 ↩︎
- Mark Andrejevic, www⁄Operational Listening (Eavesdropping), recorded on August 10, 2018 ↩︎
- ⦚bib:fac6c1a2-946f-43c4-83f5-e54fd7185c18not found For a good introduction to adversarialism, see ⦚bib:bc39dd7f-1dcc-46dc-9f52-6a16b913ff5anot found ↩︎