SeedASROptions
Options & Configuration · AiDotNet.SpeechRecognition.LLMIntegrated
Options for Seed-ASR: the LUISE Conformer encoder, the frame-splicing converter and a decoder-only LLM.
Inheritance: ModelOptions → SeedASROptions
For Beginners
The paper fixes LUISE’s depth (32 Conformer blocks), its size (nearly 2 billion parameters) and its 40 ms frame rate over 80-band filterbanks, and the converter’s four-frame splicing (160 ms per speech token). It does not publish LUISE’s width or the LLM (a mixture-of-experts model of over ten billion parameters) or its vocabulary: the width defaults to 1,600, which gives nearly 2 billion parameters over 32 blocks, and the LLM to a dense LLaMA-style decoder of 32 layers at width 4,096 with grouped-query attention.
Constructors
| Constructor | Summary |
|---|---|
SeedASROptions() |
Initializes a new instance with default values. |
SeedASROptions(SeedASROptions 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. |
SpeechLlmTrainingScope TrainingScope { get; set; } |
What training updates: every parameter by default; the paper’s SFT trains the encoder and the converter with the LLM fixed (EncoderAdapterAndLora; Seed-ASR has no LoRA). |
double LearningRate { get; set; } |
AdamW’s learning rate (the paper states no optimizer). |
double LlmRmsEpsilon { get; set; } |
Gets or sets the RMS-normalization stability constant for the language-model. |
double LlmRopeTheta { get; set; } |
Gets or sets the rotary positional-embedding frequency base for the language-model. |
double WeightDecay { get; set; } |
AdamW’s weight decay. |
int ConvKernel { get; set; } |
LUISE’s depthwise convolution kernel (the Conformer’s 31). |
int EncoderDim { get; set; } |
LUISE’s width; unpublished, 1,600 gives nearly 2 billion parameters over 32 blocks. |
int FeedForwardMultiplier { get; set; } |
LUISE’s feed-forward expansion (the Conformer’s 4). |
int LlmDim { get; set; } |
The LLM’s width (unpublished; 4,096). |
int LlmFeedForwardDim { get; set; } |
The LLM’s SwiGLU width (unpublished; 14,336). |
int MaxAudioLengthSeconds { get; set; } |
Gets or sets the longest clip, in seconds, of the reference configuration. |
int MaxTextLength { get; set; } |
The longest transcript generated, in tokens. |
int NumAttentionHeads { get; set; } |
LUISE’s attention heads (unpublished; 16). |
int NumEncoderLayers { get; set; } |
LUISE’s Conformer blocks (32, §3.2). |
int NumLlmAttentionHeads { get; set; } |
The LLM’s query heads (unpublished; 32). |
int NumLlmKvHeads { get; set; } |
The LLM’s key-value heads (unpublished; 8). |
int NumLlmLayers { get; set; } |
The LLM’s layers (unpublished; 32). |
int NumMels { get; set; } |
Mel filterbank bands (80). |
int SampleRate { get; set; } |
Gets or sets the sample rate, in Hz, that input audio is expected at. |
int SpliceFrames { get; set; } |
Consecutive 40 ms frames the converter splices into one speech token (4, §3.3). |
int VocabSize { get; set; } |
The LLM’s vocabulary (unpublished; 32,000). |
int[] Context { get; set; } |
Context tokens placed before the speech (§3.4’s context SFT); none for plain recognition. |
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 token ids to text; without one a token is its code point. |