PANNsModel<T>

Models & Types · AiDotNet.Audio.Fingerprinting

PANNs (Pretrained Audio Neural Networks) audio classifier — a CNN14-style convolutional backbone over log-mel spectrograms, trained for AudioSet tagging (Kong et al. 2020).

Inheritance: NeuralNetworkBase<T> → AudioNeuralNetworkBase<T> → PANNsModel<T>

Implements: IAudioFingerprinter<T>

How It Works

PANNs CNN14 is the de-facto baseline transfer-learning model for audio classification: four conv stages (64 → 128 → 256 → 512 channels), each stage = Conv(3×3) + BN + ReLU + Conv(3×3) + BN + ReLU + AvgPool(2×2), followed by a global average pool, a 2048-d embedding head, and a 527- class linear classifier with sigmoid activation for multi-label output. Trained on AudioSet (2 M weakly-labelled clips) at 32 kHz / 64-mel.

Reference: Kong, Q., Cao, Y., Iqbal, T., Wang, Y., Wang, W., & Plumbley, M. D. (2020), “PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition”, IEEE/ACM TASLP.

Constructors

Constructor Summary
PANNsModel(NeuralNetworkArchitecture<T> architecture, PANNsModelOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) Initializes PANNs in native training / inference mode.
PANNsModel(NeuralNetworkArchitecture<T> architecture, string modelPath, PANNsModelOptions options) Initializes PANNs in ONNX inference mode.

Properties

Property Summary
double MaxGradNormValue { get; }
int FingerprintLength { get; }
string Name { get; }

Methods

Method Summary
AudioFingerprint<T> Fingerprint(Tensor<T> audio)
AudioFingerprint<T> Fingerprint(Vector<T> audio)
Dictionary<string, double> Classify(Tensor<T> audio, double threshold) Classifies audio into AudioSet categories.
IReadOnlyList<FingerprintMatch> FindMatches(AudioFingerprint<T> query, AudioFingerprint<T> reference, int minMatchLength)
List<ValueTuple<string, double>> GetTopK(Tensor<T> audio, int k) Returns the top-K class predictions sorted descending.
ModelMetadata<T> GetModelMetadata()
Tensor<T> ForwardForTraining(Tensor<T> input) The training forward MUST run over the same log-mel spectrogram front end as inference (PredictCore), not the raw waveform.
double ComputeSimilarity(AudioFingerprint<T> fp1, AudioFingerprint<T> fp2)