CoMoSpeechOptions
Options & Configuration · AiDotNet.TextToSpeech.FlowDiffusion
Options for CoMoSpeech (Ye et al. 2023): Grad-TTS’s encoder with an EDM-preconditioned U-Net teacher, consistency-distilled into one-step synthesis.
Inheritance: ModelOptions → TtsModelOptions → EndToEndTtsOptions → CoMoSpeechOptions
For Beginners
These options configure the CoMoSpeech model. Default values follow the original paper settings.
How It Works
Defaults are the paper’s TTS configuration (§3, §4.1.2) with the values it leaves to its reference implementation
(zhenye234/CoMoSpeech, params.py, model/como.py): Grad-TTS’s encoder (width 192, 6 relative-position
layers, 2 heads, window 4, filter 768, kernel 3, dropout 0.1, pre-net of 3 kernel-5 convolutions) and duration
predictor (filter 256); Grad-TTS’s U-Net (base width 64, multipliers 1, 2, 4, time embedding scale 1000) as F_θ;
EDM with σ_data = 0.5, ε = σ_min = 0.002, σ_max = 80, ρ = 7, training σ from ln σ ~ N(−1.2, 1.2²); 50 teacher Euler
steps; consistency distillation over 50 discretization points with an EMA target at μ = 0.95; Adam at 1e-4 with
encoder and decoder gradients each clipped to norm 1; 2-second training segments; 22.05 kHz, 1024-point STFT,
hop 256, 80 mel bins.
Constructors
| Constructor | Summary |
|---|---|
CoMoSpeechOptions() |
Creates the paper’s configuration. |
CoMoSpeechOptions(CoMoSpeechOptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
double EmaDecay { get; set; } |
Gets or sets the EMA momentum of the distillation target θ⁻ (0.95, Eq. |
double LengthScale { get; set; } |
Gets or sets the duration multiplier at inference (1). |
double LogSigmaMean { get; set; } |
Gets or sets the mean of ln σ during teacher training (−1.2). |
double LogSigmaStd { get; set; } |
Gets or sets the standard deviation of ln σ during teacher training (1.2). |
double PrenetDropout { get; set; } |
Gets or sets the pre-net’s dropout (0.5). |
double Rho { get; set; } |
Gets or sets ρ of the noise-level discretization (7). |
double SigmaData { get; set; } |
Gets or sets σ_data (0.5, Eq. |
double SigmaMax { get; set; } |
Gets or sets σ_max, the largest noise level (80). |
double SigmaMin { get; set; } |
Gets or sets ε = σ_min, the smallest noise level (0.002, Eq. |
double TimePositionScale { get; set; } |
Gets or sets the scale of the noise level c_noise = ln σ / 4 in the sinusoidal embedding (1000). |
int DistillationSteps { get; set; } |
Gets or sets the discretization points of consistency distillation (50). |
int DurationPredictorFilterChannels { get; set; } |
Gets or sets the duration predictor’s filter (256). |
int EncoderKernelSize { get; set; } |
Gets or sets the encoder feed-forward and duration predictor kernel (3). |
int FlowDim { get; set; } |
Gets or sets the U-Net’s base width (64). |
int PrenetKernelSize { get; set; } |
Gets or sets the pre-net’s kernel (5). |
int PrenetLayers { get; set; } |
Gets or sets the pre-net’s convolutions (3). |
int RelativeWindow { get; set; } |
Gets or sets the encoder attention’s maximum relative position (4). |
int SamplingSeed { get; set; } |
Gets or sets the seed of the training draws and the inference sample, for repeatable runs. |
int SegmentFrames { get; set; } |
Gets or sets the training segment length in frames (2 seconds at 22.05 kHz / 256 = 172). |
int TeacherSamplingSteps { get; set; } |
Gets or sets the teacher’s Euler steps N at synthesis (50, Table 1). |
int[] DecoderDimMultipliers { get; set; } |
Gets or sets the U-Net’s width multipliers per resolution (1, 2, 4). |