RevAIOptions

Options & Configuration · AiDotNet.SpeechRecognition.ProprietaryAPI

Options for Rev AI’s Reverb ASR.

Inheritance: ModelOptions → RevAIOptions

For Beginners

These options set the size of the Conformer encoder, the feature front end and the training augmentation.

How It Works

The defaults: 18 Conformer blocks, a CTC layer and a bidirectional attention decoder of three left-to-right and three right-to-left layers, about 280M parameters (Reverb report, section 2.2). The report gives the layer counts and the total; the width (640, 8 heads, feed-forward 2,560) is the one that totals about 280M with those layers, and the kernel, CTC weight, reverse weight and label smoothing are WeNet’s U2++ defaults.

Constructors

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

Properties

Property Summary
OnnxModelOptions OnnxOptions { get; set; } ONNX runtime options.
double CtcWeight { get; set; } Weight w of the CTC loss in the joint objective w L_CTC + (1 - w) L_attention.
double LabelSmoothing { get; set; } Label smoothing of the decoder’s cross-entropy.
double ReverseWeight { get; set; } The right-to-left decoder’s share of the attention loss (0.3, WeNet U2++’s reverse_weight).
double RopeTheta { get; set; } Rotary base of the attention’s relative position encoding, standing in for WeNet’s relative positional encoding.
double TimeMaskRatio { get; set; } Largest time mask as a fraction of the utterance (section 3.1: pS = 0.05).
int AttentionDecoderFeedForward { get; set; } Feed-forward units of each decoder layer (2,560).
int AttentionDecoderLayers { get; set; } Left-to-right decoder layers (3, section 2.2).
int ConvKernel { get; set; } Depthwise convolution kernel (15, WeNet’s cnn_module_kernel).
int EncoderDim { get; set; } Conformer width (640).
int FeedForwardMultiplier { get; set; } Expansion of each half-step feed-forward module (section 2.3: 4).
int FrequencyMaskWidth { get; set; } Largest frequency mask, in mel bins (section 3.1: F = 27).
int FrequencyMasks { get; set; } SpecAugment frequency masks (section 3.1: 2).
int MaxAudioLengthSeconds { get; set; } Longest audio a transcription covers, in seconds; longer input is truncated.
int MaxTextLength { get; set; } Most tokens a transcription decodes.
int NumAttentionHeads { get; set; } Attention heads per block.
int NumEncoderLayers { get; set; } Conformer blocks (18, section 2.2).
int NumMels { get; set; } Log-mel bins (section 3.1: 80-channel filterbanks over 25 ms windows with a 10 ms stride).
int ReverseDecoderLayers { get; set; } Right-to-left decoder layers (3, section 2.2).
int SampleRate { get; set; } Audio sample rate in Hz.
int TimeMasks { get; set; } SpecAugment time masks (section 3.1: 10).
int VocabSize { get; set; } Output vocabulary size, blank (token 0) included.
string Language { get; set; } Language reported for transcriptions.
string ModelPath { get; set; } Path to an exported ONNX model, when one is used.
string[] Vocabulary { get; set; } The text of each token; tokens past the list decode as Unicode code points, and tokens in angle brackets decode to nothing.