RobustConformerOptions

Options & Configuration · AiDotNet.SpeechRecognition.Robust

Options for Wang et al. 2021’s noise-robust wav2vec 2.0 (RobustConformer).

Inheritance: ModelOptions → RobustConformerOptions

For Beginners

These options set the size of the speech encoder, the reconstruction module used only in noisy pre-training, and the regularization used while fine-tuning on transcripts.

How It Works

The defaults are section 3.2’s LibriSpeech model: wav2vec 2.0 BASE (12 Transformer blocks of width 768) as released in wav2vec_small.pt, the reconstruction module of section 2.2 weighted by λ2 = 0.1 (Equation 4), and section 3.3’s CTC fine-tuning over characters.

Constructors

Constructor Summary
RobustConformerOptions() Initializes a new instance with default values.
RobustConformerOptions(RobustConformerOptions other) Initializes a new instance by copying from another instance.

Properties

Property Summary
OnnxModelOptions OnnxOptions { get; set; } ONNX runtime options for the ONNX constructor.
bool FreezeFeatureEncoder { get; set; } Whether fine-tuning keeps the convolutional feature encoder frozen.
double ActivationDropout { get; set; } Dropout inside each feed-forward block.
double DiversityWeight { get; set; } λ1, the diversity loss’s weight in Equation 4 (0.1).
double DropoutRate { get; set; } Dropout on the Transformer’s hidden states and attention weights.
double FeatureProjectionDropout { get; set; } Dropout after the feature projection.
double FinalDropout { get; set; } Dropout before the CTC head.
double LayerDrop { get; set; } Probability of skipping each Transformer block during training.
double MaskTimeProbability { get; set; } Time-mask probability while fine-tuning.
double ReconstructionWeight { get; set; } λ2, the reconstruction loss’s weight in Equation 4 (0.1).
int EncoderDim { get; set; } Transformer width (section 3.2: 768).
int FeedForwardDim { get; set; } Feed-forward width of each block (BASE: 3,072).
int MaxAudioLengthSeconds { get; set; } Longest audio, in seconds, a transcription reads; longer input is truncated.
int MaxTextLength { get; set; } Most tokens a transcription emits.
int NumAttentionHeads { get; set; } Attention heads per block: 12, as in the released wav2vec_small.pt the paper continues from (section 3.2’s text says 8, which does not divide that checkpoint’s attention).
int NumEncoderLayers { get; set; } Transformer blocks (section 3.2: 12).
int PositionalConvGroups { get; set; } Groups of the convolutional positional embedding (16).
int ReconstructionHidden { get; set; } LSTM units per direction of the reconstruction module’s BLSTM layers; the paper does not state them, so the default makes the BLSTM output the encoder width (384 per direction).
int SampleRate { get; set; } Audio sample rate in Hz.
int VocabSize { get; set; } CTC vocabulary size, blank (token 0) included; tokens above 0 decode as Unicode code points.
string Language { get; set; } Language reported for a transcription when none is given.
string ModelPath { get; set; } Path of an exported ONNX model, for the ONNX constructor.