Dorota Orzechowska (Bio)

Echo, AI and Machine Listening

Machine listening, understood as the computational analysis of audio through which AI systems detect, classify and attribute sound, begins with an act of exclusion. Before a system can recognise a speaker, suppress noise, separate a voice, or classify a sonic scene, it must decide what counts as the sound to be heard and what may be discarded as interference. This paper argues that this decision is the constitutive operation of machine listening. The system does not find a source so much as produce one by imposing a division between signal and noise.

The argument puts pressure on Shannon’s signal/noise distinction and reads it alongside Jonathan Sterne’s account of the engineered inaudible in MP3 compression, Nina Sun Eidsheim’s formulation of the acousmatic question, and the acousmatic tradition of Pierre Schaeffer, Michel Chion and Brian Kane. This line of questioning allows machine listening to be understood as a technical and symbolic operation through which sound is rendered attributable. I develop this claim through the figure of Echo, taken as a problem of attribution. Echo is voice without a secure body, repetition without stable origin, sound shaped by reflection rather than by a single issuing source.

Machine listening answers this uncertainty by forcing acoustic events into determinate classes. In speaker identification, source separation and voice activity detection, the system must foreclose the acousmatic question and project an artificial origin, even where the sound is overlapping, partial, reverberant, or otherwise resistant to clean attribution. Where Eidsheim’s question of “who is this?” still presumes that an identity may be recovered from the voice, Echo names a prior condition in which no single origin is available for recovery at all. If sound must be given a source before it can be heard, what kind of listening has already taken place?