CIFDecoderOptions
Options & Configuration · AiDotNet.SpeechRecognition.CTCVariants
Options for the CIF-based recognizer of Dong and Xu (2020).
Inheritance: ModelOptions → CIFDecoderOptions
For Beginners
Defaults are the paper’s Mandarin (AISHELL-2) model: 40 filterbanks with deltas, a 64-filter
convolutional front-end, five self-attention layers at each of three time resolutions of width 640, a two-layer
non-autoregressive decoder and 5,230 output labels. For the LibriSpeech base model set EncoderDim 512,
FeedForwardDim 2048, PredictorWindow 5, NumDecoderLayers 3,
CtcWeight 0.25 and VocabSize 3725.
Constructors
| Constructor | Summary |
|---|---|
CIFDecoderOptions() |
Initializes a new instance with default values. |
CIFDecoderOptions(CIFDecoderOptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
OnnxModelOptions OnnxOptions { get; set; } |
Gets or sets the ONNX Runtime settings used when the model is loaded from an ONNX file. |
double CtcWeight { get; set; } |
λ1, the encoder CTC loss’s weight (0.5 for Mandarin, 0.25 for LibriSpeech). |
double LabelSmoothing { get; set; } |
The uniform label smoothing of the cross-entropy (0.2). |
double LearningRate { get; set; } |
The AdamW learning rate used when no optimizer is injected (2e-4; the paper states no optimizer). |
double QuantityWeight { get; set; } |
λ2, the quantity loss’s weight (1). |
double TailThreshold { get; set; } |
The residual weight above which inference fires once more at the tail (0.5). |
int EncoderDim { get; set; } |
The model width d (640). |
int FeedForwardDim { get; set; } |
The self-attention layers’ feed-forward width (2,560). |
int MaxAudioLengthSeconds { get; set; } |
Gets or sets the longest clip, in seconds, of the reference configuration. |
int MaxTextLength { get; set; } |
Gets or sets the maximum number of text tokens. |
int NumAttentionHeads { get; set; } |
Gets or sets the number of attention heads. |
int NumDecoderLayers { get; set; } |
Self-attention layers in the non-autoregressive decoder (2). |
int NumFilters { get; set; } |
The front-end convolutions’ filters (64). |
int NumLayersPerResolution { get; set; } |
Self-attention layers at each temporal resolution of the encoder (n = 5). |
int NumMels { get; set; } |
Filterbank bands; deltas and delta-deltas are added as channels (40). |
int NumResolutions { get; set; } |
Temporal resolutions of the encoder, separated by pooling of width 2 (3). |
int PredictorWindow { get; set; } |
The weight predictor’s convolution window (3). |
int SampleRate { get; set; } |
Gets or sets the sample rate, in Hz, that input audio is expected at. |
int VocabSize { get; set; } |
Output labels, blank (0), end of sentence (1) and pad (2) included (5,230). |
string Language { get; set; } |
Gets or sets the language code used for speech processing. |
string ModelPath { get; set; } |
Gets or sets the path of the pretrained ONNX model. |
string[] Vocabulary { get; set; } |
Optional vocabulary mapping label IDs to text. |