WavLMASR<T>

Models & Types · AiDotNet.SpeechRecognition.Foundation

WavLM ASR: speech pre-training with denoising and attention

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

Implements: ISpeechRecognizer<T>

For Beginners

WavLM extends HuBERT with: (1) gated relative position bias in self-attention for better sequence modeling; (2) denoising pre-training using overlapping utterances to learn robust representations in noisy conditions. The model is pre-trained on 94…

How It Works

References:

  • Paper: “WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing” (Chen et al., Microsoft, 2022)

WavLM extends HuBERT with: (1) gated relative position bias in self-attention for better sequence modeling; (2) denoising pre-training using overlapping utterances to learn robust representations in noisy conditions. The model is pre-trained on 94k hours of speech data. WavLM achieves state-of-the-art on the SUPERB benchmark across ASR, speaker verification, separation, and other tasks.

Example

// Create a WavLM ASR model for denoising-augmented speech pre-training
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.OneDimensional,
    taskType: NeuralNetworkTaskType.SpeechRecognition,
    inputHeight: 16000, inputWidth: 1, inputDepth: 1, outputSize: 5000);
var model = new WavLMASR<double>(architecture);
            
// Or load a pre-trained ONNX model for robust speech recognition
var onnxModel = new WavLMASR<double>(architecture, "wavlmasr.onnx");

Constructors

Constructor Summary
WavLMASR(NeuralNetworkArchitecture<T> architecture, WavLMASROptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
WavLMASR(NeuralNetworkArchitecture<T> architecture, string modelPath, WavLMASROptions options)

Properties

Property Summary
IReadOnlyList<string> SupportedLanguages { get; }
bool SupportsStreaming { get; }
bool SupportsWordTimestamps { get; }

Methods

Method Summary
IReadOnlyDictionary<string, T> DetectLanguageProbabilities(Tensor<T> audio)
IStreamingTranscriptionSession<T> StartStreamingSession(string language)
ModelMetadata<T> GetModelMetadata()
Task<TranscriptionResult<T>> TranscribeAsync(Tensor<T> audio, string language, bool? includeTimestamps, CancellationToken cancellationToken)
Tensor<T> ForwardForTraining(Tensor<T> input)
TranscriptionResult<T> Transcribe(Tensor<T> audio, string language, bool? includeTimestamps) Transcribes audio using WavLM’s denoised SSL encoder with CTC decoding.
string DetectLanguage(Tensor<T> audio)
void LoadTorchWeights(Func<string, int[], double[]> read, Func<string, bool> has) Loads a transformers WavLMForCTC state dict straight from source.