NemotronSpeechOptions

Options & Configuration · AiDotNet.SpeechRecognition.NeMo

Options for Nemotron Speech Streaming: a cache-aware Fast Conformer with an RNN-T decoder.

Inheritance: ModelOptions → NemotronSpeechOptions

For Beginners

Defaults are the released nvidia/nemotron-speech-streaming-en-0.6b: 128-band log-mel features (not normalized, since a stream has no whole-utterance statistics), 8x causal depthwise-separable subsampling to 80 ms frames, 24 Conformer blocks of width 1,024 with 8 heads and a causal kernel-9 convolution, each frame attending within its chunk and the 70 frames of chunks before it, and an RNN-T decoder (a two-layer 640-unit LSTM prediction network and a 640-wide joint network) over a 1,024-piece SentencePiece vocabulary.

Constructors

Constructor Summary
NemotronSpeechOptions() Initializes a new instance with default values.
NemotronSpeechOptions(NemotronSpeechOptions 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 FastEmitLambda { get; set; } FastEmit regularization of the transducer loss (0.005).
double NoamFactor { get; set; } The Noam schedule’s factor (the release’s lr: 3.0), scaled by the encoder width’s inverse square root.
double RopeTheta { get; set; } Rotary base standing in for the release’s relative positional encoding.
double TimeMaskRatio { get; set; } Gets or sets the maximum time-mask width as a fraction of the utterance.
double WeightDecay { get; set; } AdamW’s weight decay (1e-3).
int ConvKernel { get; set; } The causal depthwise convolution kernel (9).
int EncoderDim { get; set; } The encoder’s width (1,024).
int FeedForwardMultiplier { get; set; } The feed-forward expansion (4).
int FrequencyMaskWidth { get; set; } Gets or sets the maximum width of a SpecAugment frequency mask, in mel bins.
int FrequencyMasks { get; set; } Gets or sets the number of SpecAugment frequency masks.
int JointDim { get; set; } The joint network’s width (640).
int LeftContextFrames { get; set; } Encoder frames of left context (70, the release’s att_context_size[0]): each chunk attends to the ⌊70 / chunk⌋ chunks before it.
int LookaheadFrames { get; set; } Encoder frames of lookahead (13; the release supports 0, 1, 6 and 13): a chunk is this many frames plus one, so 13 is 1.12 s.
int MaxAudioLengthSeconds { get; set; } Gets or sets the longest clip, in seconds, of the reference configuration.
int NumAttentionHeads { get; set; } The encoder’s attention heads (8).
int NumEncoderLayers { get; set; } The encoder’s Conformer blocks (24).
int NumMels { get; set; } Log-mel bands (128).
int PredictionDim { get; set; } The prediction network’s LSTM width (640).
int PredictionLayers { get; set; } The prediction network’s LSTM layers (2).
int SampleRate { get; set; } Gets or sets the sample rate, in Hz, that input audio is expected at.
int SubsamplingChannels { get; set; } Channels of the subsampling convolutions (256).
int TimeMasks { get; set; } Gets or sets the number of SpecAugment time masks.
int VocabSize { get; set; } Output symbols, the blank (0) first: the 1,024 SentencePiece pieces follow as ids 1 to 1,024 (the release puts its blank last, at 1,024).
int WarmupSteps { get; set; } The Noam schedule’s warm-up steps (8,000).
string Language { get; set; } Gets or sets the language code used for speech processing.
string ModelPath { get; set; } Gets or sets the path to the model weights.
string[] Vocabulary { get; set; } Optional vocabulary mapping output ids to text in place of the release’s SentencePiece pieces.