IndexTTSOptions

Options & Configuration · AiDotNet.TextToSpeech.CodecBased

Options for IndexTTS.

Inheritance: ModelOptions → TtsModelOptions → IndexTTSOptions

For Beginners

These settings size IndexTTS’s networks, control how it picks each piece of sound, and set how it trains; the defaults reproduce the released 1.0 model.

How It Works

The defaults are IndexTTS 1.0 (Deng et al. 2025; IndexTeam/Index-TTS``config.yaml): a 1024-wide, 20-layer GPT-2 with 16 heads over 12000 BPE text tokens and 8194 mel codes (8192 DVAE codes, start 8192, stop 8193), conditioned by 32 perceiver latents pooled from a 6-block, 512-wide conformer over the prompt’s 100-band log mel (24 kHz, n_fft 1024, hop 256); a DVAE of 8192 codes, one per 1024 samples; and BigVGAN turning the GPT’s latents into 24 kHz audio, upsampling ×1024 from 1536 channels with SnakeBeta, conditioned on a 512-wide ECAPA-TDNN speaker embedding of the prompt. Version1_5 gives IndexTTS 1.5 (1280 wide, 24 layers, 20 heads).

Generation follows the release’s infer: beam search over 3 beams that samples its candidates (top-k 30, top-p 0.8, temperature 1, repetition penalty 10), up to 600 codes a segment of at most 120 text tokens; runs of more than 30 silent codes (code 52) shrink to 10. The paper and the release publish no training recipe; training follows XTTS’s GPT recipe, whose UnifiedVoice IndexTTS’s is: 0.01 × text cross-entropy + mel cross-entropy, AdamW at 5e-6 with betas (0.9, 0.96), weight decay 0.01 on weights only, the rate halved at steps 900k, 2.7M and 5.4M.

Constructors

Constructor Summary
IndexTTSOptions() Creates the IndexTTS 1.0 configuration.

Properties

Property Summary
bool Sample { get; set; } Sample the beam candidates (do_sample, true); otherwise take the best.
double Beta1 { get; set; } AdamW’s β1 (0.9).
double Beta2 { get; set; } AdamW’s β2 (0.96).
double Epsilon { get; set; } AdamW’s ε (1e-8).
double LearningRateDecay { get; set; } The factor at each milestone (0.5).
double MelLossWeight { get; set; } The mel cross-entropy’s weight (XTTS’s gpt_loss_mel_ce_weight, 1.0).
double RepetitionPenalty { get; set; } Repetition penalty (10).
double Temperature { get; set; } Temperature (1.0).
double TextLossWeight { get; set; } The text cross-entropy’s weight (XTTS’s gpt_loss_text_ce_weight, 0.01).
double TopP { get; set; } Top-p (0.8).
int Beams { get; set; } Beams (3).
int ConditionBlocks { get; set; } The conditioning conformer’s blocks (6).
int ConditionDim { get; set; } The conditioning conformer’s width (512).
int ConditionHeads { get; set; } The conditioning conformer’s (and perceiver’s) heads (8).
int ConditionLatents { get; set; } Perceiver latents the prompt is pooled into (32).
int ConditionLinearUnits { get; set; } The conditioning conformer’s feed-forward width (2048).
int DvaeCodebookDim { get; set; } The DVAE’s codebook width (512).
int DvaeHidden { get; set; } The DVAE’s first downsampling width (512; doubled per layer).
int DvaeKernel { get; set; } The DVAE’s downsampling kernel (3).
int DvaeLayers { get; } The DVAE’s stride-2 layers (2): as many halvings as take the mel frames (one per HopSize samples) to one per SamplesPerCode samples; -1 when no whole number of halvings does.
int DvaeResBlocks { get; set; } The DVAE’s residual blocks (3).
int KeptSilentCodes { get; set; } The silent codes kept in a row when shrinking (10).
int Layers { get; set; } GPT-2 layers (20).
int MaxGeneratedTokens { get; set; } The most mel codes generated a segment (max_mel_tokens, 600).
int MaxMelTokens { get; set; } The most mel codes a sequence holds (max_mel_tokens, 605).
int MaxSilentCodes { get; set; } More silent codes than this in a segment trigger the shrinking (30).
int MelCodes { get; set; } The mel-code vocabulary: the DVAE’s codes plus the start and stop codes (8194).
int PerceiverMultiple { get; set; } The perceiver’s feed-forward multiple (2).
int SamplesPerCode { get; } Samples per mel code (mel_length_compression, 1024): BigVGAN’s upsampling, the product of VocoderUpsampleRates.
int SamplingSeed { get; set; } The seed of generation’s sampling.
int SegmentTokens { get; set; } The most text tokens a segment holds (max_text_tokens_per_segment, 120).
int SilentCode { get; set; } The DVAE code of silence, whose long runs are shrunk (52).
int SpeakerEmbeddingDim { get; set; } The speaker embedding’s width (speaker_embedding_dim, 512).
int SpeakerEncoderAttentionChannels { get; set; } ECAPA-TDNN’s attentive pooling width (128).
int SpeakerEncoderScale { get; set; } ECAPA-TDNN’s Res2Net scale (8).
int SpeakerEncoderSeChannels { get; set; } ECAPA-TDNN’s squeeze-excitation width (128).
int TopK { get; set; } Top-k (30).
int TrainingSeed { get; set; } The seed of training’s dropout.
int VocoderInitialChannels { get; set; } BigVGAN’s first width (1536).
int[] LearningRateMilestones { get; set; } The steps at which the rate is multiplied by LearningRateDecay (MultiStepLR, [900000, 2700000, 5400000]).
int[] SpeakerEncoderChannels { get; set; } ECAPA-TDNN’s widths per stage ([512, 512, 512, 512, 1536]).
int[] SpeakerEncoderDilations { get; set; } ECAPA-TDNN’s dilations per stage ([1, 2, 3, 4, 1]).
int[] SpeakerEncoderKernels { get; set; } ECAPA-TDNN’s kernels per stage ([5, 3, 3, 3, 1]).
int[] VocoderResBlockKernels { get; set; } BigVGAN’s AMP block kernels ([3, 7, 11]).
int[] VocoderUpsampleKernels { get; set; } BigVGAN’s upsampling kernels ([8, 8, 4, 4, 4, 4]).
int[] VocoderUpsampleRates { get; set; } BigVGAN’s upsampling rates ([4, 4, 4, 4, 2, 2], ×1024).
int[][] VocoderResBlockDilations { get; set; } BigVGAN’s AMP block dilations ([1, 3, 5] each; the release’s AMPBlock1 takes three).
string BpeModelPath { get; set; } The release’s SentencePiece model (bpe.model), which turning text into tokens needs.

Methods

Method Summary
IndexTTSOptions Version1_5() The IndexTTS 1.5 configuration (IndexTeam/IndexTTS-1.5): a 1280-wide, 24-layer GPT with 20 heads, up to 600 text tokens and 800 mel codes, and a BigVGAN reading 1280-wide latents.