DeepVoice3Options
Options & Configuration · AiDotNet.TextToSpeech.Classic
Options for Deep Voice 3 (fully convolutional attention-based TTS), single-speaker configuration.
Inheritance: ModelOptions → TtsModelOptions → AcousticModelOptions → DeepVoice3Options
For Beginners
These options configure the DeepVoice3 model. Default values follow the original paper settings.
How It Works
Defaults are Table 4’s single-speaker column (Ping et al. 2018, App. C): 48 kHz audio, 4096-point FFT, 2400-sample window and 600-sample shift, r = 4, 80 mel bands; character embedding 256; encoder 7 layers of width 5 with 64 channels; decoder pre-net 128 and 256, 4 decoder layers of width 5, attention hidden size 128; position weight 1.0 and initial key rate 6.3 (query rate 1); converter 5 layers of width 5 with 256 channels; dropout keep probability 0.95; Adam at 0.001 with maximum gradient norm 100 and gradient clipping value 5. The decoding limits follow the reference implementation (r9y9/deepvoice3_pytorch: at most 200 steps, at least 10).
Constructors
| Constructor | Summary |
|---|---|
DeepVoice3Options() |
Creates the paper’s single-speaker configuration. |
DeepVoice3Options(DeepVoice3Options other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
double GradientClipValue { get; set; } |
Gets or sets the gradient clipping value applied to every component after the norm clip (5). |
double KeyPositionRate { get; set; } |
Gets or sets the keys’ position rate ω_key (6.3, the dataset’s output-to-input timestep ratio). |
double MaxGradientNorm { get; set; } |
Gets or sets the maximum gradient norm (100). |
double PositionWeight { get; set; } |
Gets or sets the positional encodings’ weight (1.0). |
double QueryPositionRate { get; set; } |
Gets or sets the queries’ position rate ω_query (1). |
int AttentionDim { get; set; } |
Gets or sets the attention hidden size (128). |
int ConvKernelSize { get; set; } |
Gets or sets the width of every convolution (5). |
int ConverterChannels { get; set; } |
Gets or sets the converter’s channels (256). |
int DecoderDim { get; set; } |
Gets or sets the decoder’s output mel width. |
int EmbeddingDim { get; set; } |
Gets or sets the character embedding width (256). |
int EncoderChannels { get; set; } |
Gets or sets the encoder convolution channels (64). |
int EncoderDim { get; set; } |
Gets or sets the encoder’s output key/value width, which is the character embedding width. |
int HiddenDim { get; set; } |
Gets or sets the decoder convolution width through the last decoder pre-net projection. |
int MaxDecoderSteps { get; set; } |
Gets or sets the most decoder steps at inference (200). |
int MinDecoderSteps { get; set; } |
Gets or sets the fewest decoder steps before the done prediction may stop decoding (10). |
int NumConverterLayers { get; set; } |
Gets or sets the converter’s convolution blocks (5). |
int WindowSize { get; set; } |
Gets or sets the analysis window in samples (2400). |
int[] DecoderPrenetSizes { get; set; } |
Gets or sets the decoder pre-net’s fully connected widths (128, 256); the last is the decoder width. |