TitaNet<T>

Models & Types · AiDotNet.Audio.Speaker

TitaNet speaker verification and embedding extraction model.

Inheritance: NeuralNetworkBase<T> → AudioNeuralNetworkBase<T> → SpeakerRecognitionBase<T> → TitaNet<T>

Implements: ISpeakerEmbeddingExtractor<T>, ISpeakerVerifier<T>

For Beginners

TitaNet is NVIDIA’s advanced voice fingerprinting model. It listens to someone speaking and creates a compact vector that uniquely identifies their voice. You can use it to verify a speaker’s identity or find who’s speaking among known voices. It comes in three sizes: Small (6M params), Medium (13M params), and Large (25M params).

Usage:

How It Works

TitaNet (Koluguri et al., ICASSP 2022) is NVIDIA’s speaker embedding model based on 1D depth-wise separable convolutions with Squeeze-Excitation and global context. TitaNet-L achieves 0.68% EER on VoxCeleb1-O, outperforming ECAPA-TDNN.

Constructors

Constructor Summary
TitaNet(NeuralNetworkArchitecture<T> architecture, TitaNetOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) Creates a TitaNet speaker model in native training mode.
TitaNet(NeuralNetworkArchitecture<T> architecture, string modelPath, TitaNetOptions options) Creates a TitaNet speaker model in native training mode.

Properties

Property Summary
ISpeakerEmbeddingExtractor<T> EmbeddingExtractor { get; }
T DefaultThreshold { get; }
bool IsOnnxMode { get; }
double MinimumDurationSeconds { get; }
double VerificationThreshold { get; }

Methods

Method Summary
IReadOnlyList<Tensor<T>> ExtractEmbeddings(IReadOnlyList<Tensor<T>> audioSegments)
ModelMetadata<T> GetModelMetadata()
SpeakerProfile<T> Enroll(IReadOnlyList<Tensor<T>> enrollmentAudio)
SpeakerProfile<T> Enroll(Tensor<T> enrollmentAudio)
SpeakerProfile<T> UpdateProfile(SpeakerProfile<T> existingProfile, Tensor<T> newAudio)
SpeakerVerificationResult<T> Verify(Tensor<T> audio, Tensor<T> referenceEmbedding)
SpeakerVerificationResult<T> Verify(Tensor<T> audio, Tensor<T> referenceEmbedding, T threshold)
SpeakerVerificationResult<T> VerifyWithReferenceAudio(Tensor<T> audio, Tensor<T> referenceAudio)
T ComputeScore(Tensor<T> audio, Tensor<T> referenceEmbedding)
T ComputeSimilarity(Tensor<T> embedding1, Tensor<T> embedding2)
T GetThresholdForFAR(double targetFAR)
Task<SpeakerVerificationResult<T>> VerifyAsync(Tensor<T> audio, Tensor<T> referenceEmbedding, CancellationToken cancellationToken)
Task<Tensor<T>> ExtractEmbeddingAsync(Tensor<T> audio, CancellationToken cancellationToken)
Tensor<T> ExtractEmbedding(Tensor<T> audio)
Tensor<T> ForwardForTraining(Tensor<T> input)