SPIRAL<T>

Models & Types · AiDotNet.SpeechRecognition.Foundation

SPIRAL: a convolution-Transformer speech encoder pre-trained to make its representation of perturbed speech match a moving-average teacher’s representation of the clean speech, with a convolutional CTC classifier.

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

Implements: ISpeechRecognizer<T>

For Beginners

SPIRAL learns from unlabeled speech with two copies of one network. The teacher hears the clean recording; the student hears it with parts blanked out, and learns to produce what the teacher produced. The teacher is a slowly moving average of the student, so the two improve together. Afterwards the student’s encoder is fine-tuned on transcribed speech.

How It Works

References:

  • Paper: “SPIRAL: Self-supervised Perturbation-Invariant Representation Learning for Speech Pre-Training” (Huang et al., ICLR 2022)

The network is the BASE model of Table 1: 128-bin log-mel input, two blocks of down-sampling convolutions and a pre-norm Transformer with convolutional relative position encoding (40 ms, then 80 ms frames), a linear projection head and the student’s convolutional predictor (section 3.2). Pretrain runs the pre-training of section 3.1: adaptive SpecAugment on the student’s input (section 3.3), random padding of the teacher’s input against positional collapse, the in-utterance contrastive loss of Equation 2, and the teacher’s moving-average update on a cosine schedule (Equation 1, section 4.2). Ordinary training fine-tunes the encoder with the convolutional classifier and CTC (section 3.5).

Example

// Create a SPIRAL recognizer
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.OneDimensional,
    taskType: NeuralNetworkTaskType.SpeechRecognition,
    inputHeight: 16000, inputWidth: 1, inputDepth: 1, outputSize: 1024);
var model = new SPIRAL<double>(architecture);
            
// Or load an exported ONNX model
var onnxModel = new SPIRAL<double>(architecture, "spiral.onnx");

Constructors

Constructor Summary
SPIRAL(NeuralNetworkArchitecture<T> architecture, SPIRALOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
SPIRAL(NeuralNetworkArchitecture<T> architecture, string modelPath, SPIRALOptions 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()
T Pretrain(Tensor<T> audio) One SPIRAL pre-training step (section 3.1) on waveforms [S] or [batch, S]: a gradient step on the student, then the teacher’s moving-average update with α on the cosine schedule from TeacherMomentumStart to TeacherMomentumEnd over Pretr…
T PretrainingLoss(Tensor<T> audio, Random random) The SPIRAL pre-training loss of waveforms [S] or [batch, S] (Huang et al.
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: the encoder, projection and CTC layer give per-frame distributions, decoded greedily.
string DetectLanguage(Tensor<T> audio)