UniAudioOptions

Options & Configuration · AiDotNet.TextToSpeech.CodecBased

Options for UniAudio’s universal codec and multi-scale Transformer.

Inheritance: ModelOptions → TtsModelOptions → CodecTtsOptions → UniAudioOptions

How It Works

Defaults follow Yang et al. 2024, Tables 8 and 18: three 1024-way RVQ codes at 50 Hz, a 24-layer global and 8-layer local causal Transformer with 12 heads, top-30 sampling at temperature 0.8, and a three-second speaker prompt. OfficialV1 selects the different axes in the released checkpoint’s config.yaml.

Constructors

Constructor Summary
UniAudioOptions() Creates the paper configuration.
UniAudioOptions(UniAudioOptions other) Initializes an independent copy of another UniAudio configuration.

Properties

Property Summary
UniAudioCodecTrainingOptions CodecTraining { get; set; } The universal codec’s reconstruction, discriminator and optimizer recipe.
UniAudioTrainingComponent TrainingComponent { get; set; } Trains the language model or independently pretrains the universal codec.
bool IncludeCodecDiscriminators { get; set; } Retains the codec critics when reconstructing a model after switching back to LM training.
bool PrefixLanguageModel { get; set; } Allows bidirectional condition attention when enabled.
bool UseReleasedTextFrontend { get; set; } Uses the released Kaldi lexicon and alignment-phone inventory for raw text.
double CodecCodebookDecay { get; set; } Gets or sets the codec codebook EMA decay (0.99).
double MaximumLengthRatio { get; set; } Gets or sets the release inference maximum-length ratio (7).
double MinimumLengthRatio { get; set; } Gets or sets the release inference minimum-length ratio (0.5).
double PromptSeconds { get; set; } Gets or sets the speaker-prompt duration (three seconds).
double Temperature { get; set; } Gets or sets sampling temperature (0.8).
int AudioEndToken { get; set; } Gets or sets the released audio-end token id.
int AudioPromptEndToken { get; set; } Gets or sets the released audio-prompt-end token id.
int AudioPromptStartToken { get; set; } Gets or sets the released audio-prompt-start token id.
int AudioStartToken { get; set; } Gets or sets the released audio-start token id.
int AudioTokenOffset { get; set; } Gets or sets the first joint-vocabulary id assigned to codec tokens.
int CodecDeadCodeThreshold { get; set; } Gets or sets the EMA dead-code threshold (2).
int CodecDimension { get; set; } Gets or sets the universal codec latent width (128).
int CodecFilters { get; set; } Gets or sets the universal codec’s base SEANet width (32).
int CodecFrameRate { get; set; } Gets the native codec’s frames per second; setting it configures the hop and codec strides.
int CodecKMeansIterations { get; set; } Gets or sets first-batch k-means iterations for a newly trained codec (50).
int CodecTotalQuantizers { get; set; } Gets or sets the codec’s trained RVQ stages (12 at 6 kbps; inference selects the first three).
int ContinuousToken { get; set; } Gets or sets the released continuous-placeholder token id.
int EndToken { get; set; } Gets or sets the released sequence-end token id.
int GlobalFeedForwardDim { get; set; } Gets or sets the global Transformer’s feed-forward width (6,144).
int GlobalStartToken { get; set; } Gets or sets the released global-transformer start token id.
int HopSize { get; set; } Gets the product of the native codec strides.
int JointVocabularySize { get; set; } Gets or sets the paper’s joint vocabulary size (4,212, including special tokens).
int LocalFeedForwardDim { get; set; } Gets or sets the local Transformer’s feed-forward width (6,144).
int LocalStartToken { get; set; } Gets or sets the released local-transformer start token id.
int MaxContextPatches { get; set; } Gets or sets the global context limit in three-token patches (3,000).
int MaximumGeneratedFrames { get; set; } Gets or sets a hard cap on newly generated audio frames.
int MinimumGeneratedFrames { get; set; } Gets or sets the minimum number of newly generated audio frames.
int PadToken { get; set; } Gets or sets the released padding token id.
int PhoneEndToken { get; set; } Gets or sets the released phoneme-end token id.
int PhoneStartToken { get; set; } Gets or sets the released phoneme-start token id.
int PhoneTokenOffset { get; set; } Gets or sets the first joint-vocabulary id assigned to phonemes.
int PhoneVocabularySize { get; set; } Gets or sets the reserved phoneme vocabulary (the release reserves 500 entries).
int SamplingSeed { get; set; } Gets or sets deterministic sampling’s seed.
int StartToken { get; set; } Gets or sets the released sequence-start token id.
int TopK { get; set; } Gets or sets top-k sampling’s candidate count (30).
int TtsTaskToken { get; set; } Gets or sets the released TTS-task token id.
int[] CodecRatios { get; set; } Gets or sets the decoder-order codec strides (8, 5, 4, 2).

Methods

Method Summary
UniAudioOptions OfficialV1() The original ep5-iter340000 checkpoint and universal codec configuration.