OuteTTSOptions

Options & Configuration · AiDotNet.TextToSpeech.Latest

Options for OuteTTS.

Inheritance: ModelOptions → TtsModelOptions → OuteTTSOptions

For Beginners

These settings size OuteTTS’s language model, its audio codec and the Whisper model that times a speaker’s words; the defaults reproduce the released 1B model.

How It Works

OuteTTS has no paper: the defaults are release 1.0 (edwko/OuteTTS and OuteAI/Llama-OuteTTS-1.0-1B’s config.json and generation_config.json): Llama 3.2 1B (width 2048, 16 layers, 32 heads over 8 key-value heads, SwiGLU 8192, Llama 3 rotary scaling, tied embeddings, 134,400 tokens) speaking DAC codes (ibm-research/DAC.speech.v1.0 at 24 kHz, 1.5 kbps: two codebooks of 1024 at 75 frames per second), sampled at temperature 0.4 with a repetition penalty of 1.1 over the last 64 tokens, top-k 40, top-p 0.9 and min-p 0.05, in chunks of 10–30 words. Speaker profiles time their words with Whisper large-v3-turbo. Qwen3 gives OuteTTS-1.0-0.6B.

Constructors

Constructor Summary
OuteTTSOptions() Creates the Llama-OuteTTS-1.0-1B configuration.
OuteTTSOptions(OuteTTSOptions other) Copies another instance.

Properties

Property Summary
OuteTTSBackbone Backbone { get; set; } The language model (and with it the tokenizer).
bool Chunked { get; set; } Whether long text is generated in chunks (the release’s default) rather than in one pass.
bool QueryKeyNorm { get; set; } Qwen3’s per-head query and key RMSNorm (off for Llama).
double Beta1 { get; set; } AdamW β₁ (0.9).
double Beta2 { get; set; } AdamW β₂ (0.999).
double Epsilon { get; set; } AdamW ε (1e-8).
double MaxGradNorm { get; set; } The gradient norm clipped to (1.0).
double MaxSpeakerSeconds { get; set; } The longest speaker recording, in seconds (20).
double MinP { get; set; } Min-p (0.05).
double RepetitionPenalty { get; set; } Repetition penalty (1.1).
double RmsEpsilon { get; set; } RMSNorm ε (1e-5).
double RopeHighFrequencyFactor { get; set; } Llama 3’s high-frequency factor (4).
double RopeLowFrequencyFactor { get; set; } Llama 3’s low-frequency factor (1).
double RopeScalingFactor { get; set; } Llama 3’s rotary scaling factor (32; 0 for none).
double RopeTheta { get; set; } Rotary base (500000).
double Temperature { get; set; } Sampling temperature (0.4; 0 samples greedily).
double TopP { get; set; } Top-p (0.9).
double[] AlignerTemperatures { get; set; } The temperatures Whisper falls back through when transcribing a speaker (0, 0.2, …, 1).
int AlignerDecoderLayers { get; set; } The Whisper aligner’s decoder layers (4).
int AlignerEncoderLayers { get; set; } The Whisper aligner’s encoder layers (32).
int AlignerHeads { get; set; } The Whisper aligner’s heads (20).
int AlignerMels { get; set; } The Whisper aligner’s mel bins (128, large-v3-turbo).
int AlignerWidth { get; set; } The Whisper aligner’s width (1280).
int ChunkMaxWords { get; set; } Most words in a chunk (30).
int ChunkMinWords { get; set; } Fewest words in a chunk (10).
int CodecCodebookDim { get; set; } Codebook width (8).
int CodecCodebookSize { get; set; } Entries per codebook (1024).
int CodecCodebooks { get; set; } Codebooks the language model speaks (2).
int CodecDecoderDim { get; set; } The codec decoder’s first width (1536).
int CodecEncoderDim { get; set; } The codec encoder’s first width (64).
int CodecLatentDim { get; set; } The codec latent width (1024).
int HeadDim { get; set; } Head width (64; 0 for width / heads).
int Heads { get; set; } Attention heads (32).
int Intermediate { get; set; } SwiGLU width (8192).
int KeyValueHeads { get; set; } Key-value heads (8).
int Layers { get; set; } Decoder layers (16).
int MaxLength { get; set; } Prompt and generated tokens at most (8192).
int NumDecoderLayers { get; set; } The released causal language model’s decoder blocks.
int NumHeads { get; set; } The released causal language model’s query attention heads.
int RepetitionRange { get; set; } Tokens the repetition penalty looks back over (64).
int RopeOriginalMaxPositions { get; set; } Llama 3’s original context (8192).
int SamplingSeed { get; set; } Seed of sampling and of the speaker transcriptions’ fallback draws.
int TopK { get; set; } Top-k (40).
int TrainingSteps { get; set; } Updates the learning rate decays linearly to 0 over (0: constant).
int[] CodecDecoderRates { get; set; } The codec decoder’s upsampling (8, 5, 4, 2).
int[] CodecEncoderRates { get; set; } The codec encoder’s strides (2, 4, 5, 8: a hop of 320, 75 frames per second at 24 kHz).

Methods

Method Summary
OuteTTSOptions Qwen3() OuteTTS-1.0-0.6B: Qwen3 0.6B (width 1024, 28 layers, 16 heads of width 128 over 8 key-value heads, SwiGLU 3072, per-head query and key RMSNorm, rotary base 10⁶, 157,760 tokens).