Qwen3ASRSmallOptions

Options & Configuration · AiDotNet.SpeechRecognition.AlibabaASR

Options for Qwen3-ASR-0.6B: an AuT audio encoder, a projector and the Qwen3-0.6B decoder (Qwen3-ASR technical report, 2026).

Inheritance: ModelOptions → Qwen3ASRSmallOptions

For Beginners

Defaults are the released Qwen3-ASR-0.6B configuration: 128-band Whisper log-mel features, an AuT encoder of 18 layers at width 896 (14 heads, feed-forward 3584) that emits 12.5 tokens per second, and Qwen3-0.6B (1,024 wide, 28 layers, 16 query and 8 key-value heads of width 128, SwiGLU 3,072, 151,936 tokens, tied embeddings). Fine-tuning uses AdamW at 2e-5 with a linear schedule, as the release’s SFT script does.

Constructors

Constructor Summary
Qwen3ASRSmallOptions() Initializes a new instance with default values.
Qwen3ASRSmallOptions(Qwen3ASRSmallOptions 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 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.
int ChunkFrames { get; set; } Spectrogram frames per convolution chunk (100, one second).
int DownsampleChannels { get; set; } The convolutional front end’s channels (480).
int EncoderDim { get; set; } The AuT encoder’s width (896).
int EncoderFeedForwardDim { get; set; } The AuT encoder’s feed-forward width (3584).
int LlmHeadDim { get; set; } Qwen3-0.6B’s head width (128).
int LlmHeads { get; set; } Qwen3-0.6B’s query heads (16).
int LlmHiddenDim { get; set; } Qwen3-0.6B’s width (1,024).
int LlmIntermediateDim { get; set; } Qwen3-0.6B’s SwiGLU width (3,072).
int LlmKeyValueHeads { get; set; } Qwen3-0.6B’s key-value heads (8).
int MaxAudioLengthSeconds { get; set; } The longest recording transcribed (20 minutes, the report’s single-utterance limit).
int MaxTextLength { get; set; } The longest answer generated, in tokens.
int NumAttentionHeads { get; set; } The AuT encoder’s attention heads (14).
int NumEncoderLayers { get; set; } The AuT encoder’s layers (18).
int NumLlmLayers { get; set; } Qwen3-0.6B’s layers (28).
int NumMels { get; set; } Whisper log-mel bands (128).
int SampleRate { get; set; } Gets or sets the sample rate, in Hz, that input audio is expected at.
int VocabSize { get; set; } The decoder’s vocabulary (151,936).
int WarmupSteps { get; set; } Linear warmup steps.
int WindowChunks { get; set; } Chunks per encoder attention window (8).
string Context { get; set; } Context text (hotwords, background) placed in the system message; empty by default.
string Language { get; set; } The language to force (“English”, “Chinese”, …): the prompt then ends with language {Language}<asr_text> and the answer is the transcript alone.
string ModelPath { get; set; } Gets or sets the path to the model weights.
string[] Vocabulary { get; set; } Optional vocabulary mapping token ids to text when the vocabulary is not Qwen’s.