Clasificación de escenas acústicas

Acoustic scene classification (ASC) asks a computer to name the environment a recording was made in, such as a metro station, an office, a shopping mall or a park, from its sound alone. Barchiesi, Giannoulis, Stowell and Plumbley, in a 2015 review in IEEE Signal Processing Magazine, define it simply as classifying environments from the sounds they produce. It differs from sound event detection, which locates individual sounds like a door slam or a dog bark in time. A scene label describes the whole clip, so the model must summarize the overall texture of many overlapping sounds: traffic hum, footsteps, announcements, birdsong and reverberation.

A typical modern system works in four steps. First, the audio is cut into fixed-length clips, for example 1 second in the 2024 DCASE challenge or 30 seconds in the 2016 edition. Second, each clip is converted into a time-frequency picture, usually a log-mel spectrogram, which shows how much energy falls into perceptually spaced frequency bands over time. Third, a classifier, today usually a convolutional neural network (CNN) that scans the spectrogram like an image, produces a score for each scene class. Fourth, the highest score becomes the label, sometimes after averaging over several clips to smooth out brief misleading sounds.

Earlier systems used hand-built features. The benchmark method in the 2015 review, and the baseline in the DCASE 2016 challenge, computed mel-frequency cepstral coefficients (MFCCs), compact summaries of the spectrum's shape, and modeled each scene with a Gaussian mixture model (GMM), choosing the scene whose model gave the input the highest likelihood. In the challenge the review describes, only three submitted algorithms significantly outperformed that baseline, and the best one matched the median accuracy of human listeners. Some scene pairs confused both people and machines.

The field's main milestones track the DCASE (Detection and Classification of Acoustic Scenes and Events) challenges. The 2016 acoustic scene task used the TUT Acoustic Scenes 2016 dataset: 15 scenes including bus, cafe or restaurant, car, city center, forest path, grocery store, home, library, metro station, office and urban park, recorded with binaural in-ear microphones at 44.1 kHz and 24 bits and split into 30-second segments. Its MFCC and GMM baseline reached 72.5% accuracy in development cross-validation, and the best of 48 submitted systems reached 89.7% on the evaluation set. Later editions shifted toward real-world constraints. By 2024 the task used the TAU Urban Acoustic Scenes 2022 Mobile dataset, with 1-second clips from 3 real and 6 simulated recording devices in development, extra unseen devices in evaluation, a 128 kB limit on model size and a 30 million multiply-accumulate limit per second of audio. Teams also had to train on subsets as small as 5% of the data, and the baseline's accuracy ranged from 42.40% at 5% to 56.99% with all the data.

ASC sits beside several neighboring approaches. Sound event detection gives finer detail but needs time-stamped labels that are costly to produce. General audio taggers trained on large collections, such as the PANNs models trained on AudioSet, recognize hundreds of sound types and can be fine-tuned for scenes; the PANNs authors report a mean average precision of 0.439 on AudioSet tagging and transfer to acoustic scene classification among other tasks. Location data such as GPS can say where a device is but not whether it is in a quiet car or a noisy one, which is why hearing aids rely on sound. Phonak's AutoSense OS, for example, uses classical machine learning trained on audio labeled by audiology engineers to pick among six sound classes.

Hearing aids show how a scene label is used in practice. Sonova's fact sheet describes AutoSense OS as a combination of supervised and unsupervised learning, trained on its own media database, whose output selects the hearing program best suited to the current sound, with no input from the wearer. The fact sheet names the main risk plainly: the classifier may choose a program that is not the best one, though the wearer still receives amplification. That framing is typical of scene classification in products, where a wrong label degrades comfort or clarity rather than causing outright failure, so designers trade accuracy against power, memory and response time.

The hard problems are well known from the challenges. Device mismatch is the biggest: a model trained on one microphone performs worse on another, which is why DCASE evaluates on unseen devices. Scenes overlap, so a street with a tram passing may sound like a tram, and an indoor market may sound like a shopping mall. Short clips carry less evidence, and the 2024 baseline's accuracy below 57% on 1-second clips shows how hard the setting is. Class lists are arbitrary and culture-specific, since a metro station in Helsinki may not sound like one in Mumbai. Finally, recordings of public places can capture conversations, so collecting data raises privacy questions.

Deployment costs vary with the target. On a server, a large pretrained network can classify scenes cheaply. On a hearing aid or earbud, power and memory are tight, which is exactly what DCASE's complexity limits simulate: 128 kB holds only about 128,000 parameters in 8-bit integers, and the 2024 baseline used 61,148 parameters and about 29.4 million multiply-accumulates per inference. Training a competitive model needs labeled recordings from many places and devices; the 2024 development training split alone contained 139,620 one-second segments.