UniSpeech<T>

Models & Types · AiDotNet.SpeechRecognition.Foundation

UniSpeech: unified pre-training for speech representation learning

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

Implements: ISpeechRecognizer<T>

For Beginners

UniSpeech unifies self-supervised and supervised pre-training in a single framework. The model jointly optimizes a contrastive self-supervised loss and a CTC loss when labeled data is available. This multi-task pre-training allows the model to lev…

How It Works

References:

  • Paper: “UniSpeech: Unified Pre-training for Self-Supervised Learning and Supervised Learning for ASR” (Wang et al., Microsoft, 2021)

UniSpeech unifies self-supervised and supervised pre-training in a single framework. The model jointly optimizes a contrastive self-supervised loss and a CTC loss when labeled data is available. This multi-task pre-training allows the model to leverage both unlabeled and labeled speech data during pre-training, producing representations that are more aligned with the downstream ASR task than purely self-supervised approaches.

Example

// Create a UniSpeech model for unified self-supervised and supervised pre-training
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.OneDimensional,
    taskType: NeuralNetworkTaskType.SpeechRecognition,
    inputHeight: 16000, inputWidth: 1, inputDepth: 1, outputSize: 5000);
var model = new UniSpeech<double>(architecture);
            
// Or load a pre-trained ONNX model for multi-task speech recognition
var onnxModel = new UniSpeech<double>(architecture, "unispeech.onnx");

Constructors

Constructor Summary
UniSpeech(NeuralNetworkArchitecture<T> architecture, UniSpeechOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
UniSpeech(NeuralNetworkArchitecture<T> architecture, string modelPath, UniSpeechOptions 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 UniSpeech’s unified SSL/supervised encoder with CTC.
string DetectLanguage(Tensor<T> audio)
void LoadTorchWeights(Func<string, int[], double[]> read) Loads a transformers *ForCTC state dict straight from source.