SceneClassifier<T>

Models & Types · AiDotNet.Audio.Classification

Acoustic scene classification model for identifying recording environments.

Inheritance: NeuralNetworkBase<T> → AudioNeuralNetworkBase<T> → AudioClassifierBase<T> → SceneClassifier<T>

Implements: ISceneClassifier<T>

For Beginners

Scene classification answers “Where was this recorded?”:

  • Indoor: office, home, shopping mall, restaurant, library
  • Outdoor: street, park, beach, forest, construction site
  • Transportation: bus, train, metro, airport, car

Usage with ONNX model:

Usage for training:

How It Works

Classifies audio recordings by their acoustic environment or scene context with the DCASE 2017 Task 1 baseline system (Mesaros et al. 2017, sections 2.2 and 3): log mel-band energies in 40 ms frames with 50% overlap over 40 bands from 0 to 22050 Hz, a 5-frame context (a 200-value vector), and a multilayer perceptron of two dense layers of 50 units with 20% dropout under a softmax over the 15 TUT Acoustic Scenes 2017 classes, trained with Adam at a learning rate of 0.001. A recording’s label is the majority vote of its frame decisions.

The baseline also standardizes features with training-set statistics; feed standardized context vectors to Train and Predict when your recordings’ levels vary widely.

Constructors

Constructor Summary
SceneClassifier(NeuralNetworkArchitecture<T> architecture, SceneClassifierOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) Creates a SceneClassifier for native training mode.
SceneClassifier(SceneClassifierOptions options) Creates a SceneClassifier with default options for basic classification.
SceneClassifier(string modelPath, SceneClassifierOptions options) Creates a SceneClassifier for ONNX inference mode.

Properties

Property Summary
IReadOnlyList<string> Scenes { get; } Gets the scenes (alias for SupportedScenes for legacy API compatibility).
IReadOnlyList<string> SupportedScenes { get; } Gets the list of scenes this model can classify.
double MinimumDurationSeconds { get; } Gets the minimum audio duration required for reliable classification.

Methods

Method Summary
IReadOnlyDictionary<string, T> GetSceneProbabilities(Tensor<T> audio) Gets scene probabilities for all supported scenes.
IReadOnlyList<ScenePrediction<T>> GetTopScenes(Tensor<T> audio, int k) Gets top-K scene predictions.
ModelMetadata<T> GetModelMetadata() Gets model metadata for serialization.
SceneClassificationResult<T> Classify(Tensor<T> audio) Classifies the acoustic scene of an audio recording.
SceneTrackingResult<T> TrackSceneChanges(Tensor<T> audio, double segmentDuration) Tracks scene changes over time in longer audio.
Task<SceneClassificationResult<T>> ClassifyAsync(Tensor<T> audio, CancellationToken cancellationToken) Classifies the acoustic scene asynchronously.
Tensor<T> ContextFeatures(Tensor<T> audio) The classifier input of a recording, [frames, ContextFrames * NumMels]: per frame, the log mel-band energies of the ContextFrames frames centred on it (edges repeat the first and last frame).
Tensor<T> ExtractAcousticFeatures(Tensor<T> audio) Extracts acoustic features used for scene classification.
ValueTuple<string, T> ClassifyCategory(Tensor<T> audio) Classifies audio and returns category with confidence (legacy API compatibility).

Fields

Field Summary
string[] StandardScenes Standard acoustic scene labels (DCASE-style).