ZonosOptions
Options & Configuration · AiDotNet.TextToSpeech.CodecBased
Options for Zonos.
Inheritance: ModelOptions → ZonosOptions
For Beginners
These settings size Zonos, pick its backbone, control how it samples, set the default speaking style, and set how it trains; the defaults reproduce the released transformer model.
How It Works
The defaults are Zonos-v0.1-transformer (Zyphra 2025; Zyphra/Zonos zonos/config.py, the release’s
config.json): width 2048, 26 attention layers of 16 query heads over 4 key-value heads with rotary positions on adjacent
pairs, gated feed-forward 8192, over DAC’s nine 1024-code codebooks at 44.1 kHz. Hybrid gives
Zonos-v0.1-hybrid: 46 layers, attention at 0, 4, …, 44 (rotary on halves) and Mamba2 elsewhere (state 128, head width 64,
expansion 2, one group), with four more conditioners.
Generation follows Zonos.generate with make_cond_dict’s defaults: classifier-free guidance at 2.0,
min-p 0.1, a repetition penalty of 3 over the last 2 frames, up to 30 seconds; English, a neutral-leaning emotion mix,
fmax 22050 Hz, pitch deviation 20 Hz and 15 phonemes a second. Training (no code was released) scores each head on its
next delayed code, drops each optional conditioner 10% of the time, and uses the Trainer’s default AdamW.
Constructors
| Constructor | Summary |
|---|---|
ZonosOptions() |
Creates the Zonos-v0.1-transformer configuration. |
ZonosOptions(ZonosOptions other) |
Copies another instance. |
Properties
| Property | Summary |
|---|---|
DACOptions Codec { get; set; } |
The codec (DAC 44.1 kHz, 9 codebooks); load its released weights with LoadPretrainedWeights. |
List<ZonosConditioner> Conditioners { get; set; } |
The prefix conditioners, in order (the transformer’s seven). |
List<string> UnconditionalInputs { get; set; } |
Conditioners left unconditional (make_cond_dict’s vqscore_8 and dnsmos_ovrl). |
OnnxModelOptions OnnxOptions { get; set; } |
Settings for the ONNX graph runtime. |
ZonosBackbone Backbone { get; set; } |
The backbone (the transformer). |
bool RotaryInterleaved { get; set; } |
Whether the rotary embedding turns adjacent pairs (the transformer) rather than halves (the hybrid). |
bool SpeakerNoised { get; set; } |
The hybrid’s speaker-noised flag (false). |
double Beta1 { get; set; } |
AdamW’s β₁ (0.9, Hugging Face’s default). |
double Beta2 { get; set; } |
AdamW’s β₂ (0.999, Hugging Face’s default). |
double CfgScale { get; set; } |
Classifier-free guidance’s scale (2.0; the release requires it to differ from 1). |
double ConditionerDropout { get; set; } |
In training, the chance each optional conditioner is replaced by its unconditional vector (0.1). |
double CtcLoss { get; set; } |
The hybrid’s target CTC loss (0). |
double DnsmosOverall { get; set; } |
The hybrid’s target DNSMOS (4). |
double Epsilon { get; set; } |
AdamW’s ε (1e-8, Hugging Face’s default). |
double Fmax { get; set; } |
The highest frequency in Hz (22050; use 22050 for voice cloning). |
double LearningRate { get; set; } |
The native AdamW learning rate (5e-5). |
double MaxGradNorm { get; set; } |
The gradient-norm clip (1.0, Hugging Face’s default; 0 for none). |
double MinP { get; set; } |
Min-p (0.1). |
double NormEpsilon { get; set; } |
LayerNorm’s ε (1e-5). |
double PitchStd { get; set; } |
The pitch’s standard deviation in Hz (20; 20–45 normal, 60–150 expressive). |
double RepetitionPenalty { get; set; } |
The repetition penalty (3.0; 1 for none). |
double SpeakingRate { get; set; } |
Phonemes a second (15; 10 slow, 30 very fast). |
double Temperature { get; set; } |
The temperature (1.0; 0 takes the most likely code). |
double TopP { get; set; } |
Nucleus sampling’s mass (0: off). |
double UnifiedConfidence { get; set; } |
The unified sampler’s confidence term. |
double UnifiedLinear { get; set; } |
NovelAI’s unified sampler’s linear term (0: off). |
double UnifiedQuadratic { get; set; } |
The unified sampler’s quadratic term. |
double WeightDecay { get; set; } |
The native AdamW weight decay (0). |
double[] Emotion { get; set; } |
The emotion mix: happiness, sadness, disgust, fear, surprise, anger, other, neutral (normalized to sum 1). |
double[] VqScore { get; set; } |
The hybrid’s target VQScore for each eighth of the audio (0.78). |
int AttentionIntermediate { get; set; } |
The attention layers’ gated feed-forward width (8192). |
int ChunkSize { get; set; } |
The SSD algorithm’s chunk length (256); it changes only the order of the arithmetic. |
int Codebooks { get; set; } |
DAC codebooks (9). |
int ConvWidth { get; set; } |
Mamba2’s convolution width (4). |
int EndToken { get; set; } |
The end-of-audio token (1024). |
int Expand { get; set; } |
Mamba2’s expansion (2). |
int Groups { get; set; } |
Mamba2’s groups of B and C (1). |
int Heads { get; set; } |
Query heads (16). |
int HiddenDim { get; set; } |
The backbone and conditioning width (2048). |
int KeyValueHeads { get; set; } |
Key-value heads (4). |
int Layers { get; set; } |
Layers (26; the hybrid has 46). |
int MambaHeadDim { get; set; } |
Mamba2’s head width (64). |
int MambaIntermediate { get; set; } |
The Mamba2 layers’ gated feed-forward width (0: none). |
int MaskToken { get; set; } |
The mask token of the delay pattern (1025). |
int MaxNewTokens { get; set; } |
The most frames generated (2580: 30 seconds). |
int MaxTextLength { get; set; } |
The input length advertised to TTS clients (2048). |
int RepetitionWindow { get; set; } |
The frames the repetition penalty looks back over (2). |
int RotaryDim { get; set; } |
Head dimensions the rotary embedding turns (128: all). |
int SampleRate { get; set; } |
The waveform sample rate, matching the DAC configuration (44100). |
int SamplingSeed { get; set; } |
The seed of sampling and conditioner dropout, and the initialization fallback when neither the shared Seed nor the architecture specifies a seed (0). |
int StateDim { get; set; } |
Mamba2’s state size (128). |
int TopK { get; set; } |
Top-k (0: off). |
int TrainingSteps { get; set; } |
The total optimizer steps the linear decay to 0 spans (0: a constant learning rate). |
int[] AttentionLayers { get; set; } |
The layers that attend; null selects all layers for Transformer or every fourth layer for Hybrid. |
string Language { get; set; } |
The eSpeak language (en-us); one of Languages. |
string ModelPath { get; set; } |
The exported ONNX graph path, when used. |
string SudachiDictionary { get; set; } |
For Japanese: sudachidict_full’s system.dic (null: the SUDACHI_DICT environment variable). |
string SudachiResources { get; set; } |
For Japanese: SudachiPy’s resources directory with char.def, unk.def and rewrite.def (null: the SUDACHI_RESOURCES environment variable). |
Methods
| Method | Summary |
|---|---|
ZonosOptions Hybrid() |
The Zonos-v0.1-hybrid configuration. |