Chirp3Options

Options & Configuration · AiDotNet.SpeechRecognition.Multilingual

Options for Chirp 3, built as Google USM-CTC: a BEST-RQ pre-trained Conformer with a CTC layer.

Inheritance: ModelOptions → Chirp3Options

For Beginners

These options set the size of Chirp 3’s Conformer encoder, how it learns from unlabeled speech, and how far its attention reaches.

How It Works

The defaults are the paper’s Conformer-2B (Zhang et al. 2023, Table 2: 32 layers, 1536 wide, 16 heads, kernel 5) over 128-bin log-mel features (section 2.7), with BEST-RQ pre-training under a 16-codebook multi-softmax loss and 8-second chunk-wise attention (section 2.4). EncoderDim 1024, NumEncoderLayers 24 and NumAttentionHeads 8 give Conformer-0.6.

Constructors

Constructor Summary
Chirp3Options() Initializes a new instance with default values.
Chirp3Options(Chirp3Options other) Initializes a new instance by copying from another instance.

Properties

Property Summary
OnnxModelOptions OnnxOptions { get; set; } ONNX runtime options.
double LogMelFloor { get; set; } Floor added to the mel energies before the logarithm.
double MaskNoiseStd { get; set; } Standard deviation of the zero-mean normal noise that replaces masked frames (BEST-RQ’s default: 0.1).
double MaskProbability { get; set; } Probability that each input frame starts a masked span (BEST-RQ’s default: 0.01).
double RopeTheta { get; set; } Rotary base of the attention’s relative position encoding, which stands in for the paper’s relative attention (section 2.1).
int AttentionChunkSeconds { get; set; } Chunk-wise attention window in seconds (section 2.4: an 8-second chunk was best); zero attends over the whole input.
int CodeDim { get; set; } Codebook entry dimension, the height of each projection matrix (BEST-RQ’s default: 16).
int CodebookSize { get; set; } Entries per random-projection codebook, which is also each pre-training softmax’s width (BEST-RQ’s default, which section 2.7 follows: 8192).
int ConvKernel { get; set; } Depthwise convolution kernel (Table 2: 5).
int EncoderDim { get; set; } Conformer width (Table 2, Conformer-2B: 1536; Conformer-0.6: 1024).
int FeedForwardMultiplier { get; set; } Feed-forward expansion of each half-step feed-forward module (Gulati et al.
int HopLength { get; set; } Frame stride in samples (10 ms, BEST-RQ’s, which the 4x reduction turns into 40 ms encoder frames).
int MaskLength { get; set; } Masked span length in input frames (BEST-RQ’s default: 400 ms, 40 frames at the 10 ms stride).
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 (Table 2, Conformer-2B: 16; Conformer-0.6: 8).
int NumCodebooks { get; set; } Codebooks of the multi-softmax loss, each with its own quantizer and softmax layer (sections 2.2.1 and 2.7: 16).
int NumEncoderLayers { get; set; } Conformer blocks (Table 2, Conformer-2B: 32; Conformer-0.6: 24).
int NumMels { get; set; } Log-mel bins (section 2.7: 128-dimensional log-mel filterbank coefficients).
int SampleRate { get; set; } Audio sample rate in Hz (section 2.7: audio is resampled to 16 kHz).
int StackedFrames { get; set; } Input frames stacked per quantized label, matching the encoder’s 4x reduction (BEST-RQ’s default: 4).
int VocabSize { get; set; } CTC vocabulary size, blank (token 0) included; tokens above 0 decode as Unicode code points.
int WindowLength { get; set; } Analysis window in samples (25 ms at 16 kHz; the paper does not state it).
string Language { get; set; } Language reported for transcriptions.
string ModelPath { get; set; } Path to an exported ONNX model, when one is used.