Sound Event Detection

Sound Event Detection (SED) is the task of identifying the presence of sound events in an audio recording, estimating their temporal positions (start and end times), and classifying them into predefined categories.

soundevent.evaluation.tasks.sound_event_detection

Sound event detection evaluation.

Functions:

Name	Description
evaluate_clip
evaluate_sound_event_detection	Evaluate sound event detections against ground truth annotations.
sound_event_detection

Attributes

EXAMPLE_METRICS = () module-attribute

RUN_METRICS = ((terms.mean_average_precision, metrics.mean_average_precision), (terms.balanced_accuracy, metrics.balanced_accuracy), (terms.accuracy, metrics.accuracy), (terms.top_3_accuracy, metrics.top_3_accuracy)) module-attribute

SOUNDEVENT_METRICS = ((terms.true_class_probability, metrics.true_class_probability),) module-attribute

Classes

ClipPrediction

Bases: Protocol, Generic[Detection]

Protocol defining the requirements for a clip prediction object.

Attributes:

Name	Type	Description
clip	Clip
detections	Sequence[Detection]

Attributes

clip instance-attribute

detections instance-attribute

Functions

compute_overall_metrics(true_classes, predicted_classes_scores)

Compute evaluation metrics based on true classes and predicted scores.

evaluate_clip(clip_annotations, clip_predictions, encoder)

evaluate_sound_event(sound_event_prediction, sound_event_annotation, encoder)

evaluate_sound_event_detection(clip_predictions, clip_annotations, affinity, score=None, affinity_threshold=0, strict_match=False)

Evaluate sound event detections against ground truth annotations.

This function matches predictions to annotations for each clip individually.

Parameters:

Name	Type	Description	Default
clip_predictions	Sequence[ClipPrediction[Detection]]	A sequence of prediction objects. Each object must contain a reference to the clip and a sequence of detections.	required
clip_annotations	Sequence[ClipAnnotation]	A sequence of ground truth annotations corresponding to the same clips.	required
affinity	Callable[[Detection, SoundEventAnnotation], float]	A function that computes the affinity score (e.g., IoU) between a detection and a ground truth annotation.	required
score	Callable[[Detection], float] | None	A function to extract the confidence score from a detection. Used to sort detections greedily. If None, detections are processed in the order provided.	None
affinity_threshold	float	The minimum affinity score required for a valid match. Matches with scores less than or equal to this value are discarded. Defaults to 0.0.	0
strict_match	bool	If True, a detection is only matched if its highest affinity target is available. If False (default), it falls back to the next best available target.	False

Yields:

Name	Type	Description
clip	Clip	The clip associated with the match.
match	Match[Detection, SoundEventAnnotation]	A named tuple containing the matching results, see [Match][].

Raises:

Type	Description
ValueError	If the number of predictions and annotations differs, or if the sets of clip UUIDs do not match exactly.

sound_event_detection(clip_predictions, clip_annotations, tags)