ChatterboxOptions

Options & Configuration · AiDotNet.TextToSpeech.VoiceCloning

Options for Resemble AI’s Chatterbox releases (the original English model by default; see Multilingual).

Inheritance: ModelOptions → CosyVoiceFamilyOptions → CosyVoiceConformerOptions → ChatterboxOptions

How It Works

The defaults reproduce the released 2025 model: the 520M T3 decoder (30 Llama layers, 1,024 channels), the 25 Hz S3Tokenizer-v2 vocabulary, the causal S3Gen flow, CAM++ and the 24 kHz HiFT vocoder.

Constructors

Constructor Summary
ChatterboxOptions() Initializes a new instance with default values.
ChatterboxOptions(ChatterboxOptions other) Initializes a copy of every Chatterbox setting.

Properties

Property Summary
ChatterboxVariant Variant { get; set; } The released checkpoint this model follows: the text front end and the files LoadPretrainedWeights reads.
bool Watermark { get; set; } Whether synthesis puts Resemble’s Perth implicit watermark on its output, as the release does (true).
double ClassifierFreeGuidanceWeight { get; set; } Gets or sets the weight of the classifier free guidance objective.
double Exaggeration { get; set; } Gets or sets the exaggeration.
double GenerationTopP { get; set; } Gets or sets the cumulative probability threshold for generated-token nucleus sampling.
double MinP { get; set; } Gets or sets the minimum token probability relative to the most likely token during sampling.
double PerceiverDropout { get; set; } Gets or sets the dropout probability for the perceiver.
double RepetitionPenalty { get; set; } Gets or sets the penalty applied to repeated tokens during generation.
double RmsEpsilon { get; set; } Gets or sets the RMS-normalization stability constant for the model.
double RopeHighFrequencyFactor { get; set; } Gets or sets the rotary positional embedding high frequency factor.
double RopeLowFrequencyFactor { get; set; } Gets or sets the rotary positional embedding low frequency factor.
double RopeScalingFactor { get; set; } Gets or sets the rotary positional-embedding scaling factor.
double RopeTheta { get; set; } Gets or sets the rotary positional-embedding frequency base for the model.
double SpeechLossWeight { get; set; } Gets or sets the weight of the speech loss.
double Temperature { get; set; } Gets or sets the logit temperature used during sampling.
double TextLossWeight { get; set; } Gets or sets the weight of the text loss.
double VoiceEncoderPartialRate { get; set; } Partials per second setting used by embeds_from_wavs; 1.3 gives the release’s 77-frame step.
double? ReferenceLoudness { get; set; } The integrated loudness (LUFS, ITU-R BS.1770 as pyloudnorm measures it) a reference recording is normalized to before conditioning, or null to use it as recorded (the original and multilingual releases).
int CodebookSize { get; set; } S3Tokenizer v2’s fixed eight-dimensional ternary FSQ vocabulary.
int DecoderReferenceSeconds { get; set; } Gets or sets the decoder reference duration, in seconds.
int DistillationTeacherSteps { get; set; } Euler steps of the frozen teacher S3Gen per distillation target when training a mean-flow (Turbo, Nano) decoder: the teacher’s average velocity over [t, r] (IntMeanFlow, Eq.
int EncoderReferenceSeconds { get; set; } Gets or sets the encoder reference duration, in seconds.
int GenerationTopK { get; set; } Top-k sampling of speech tokens (Hugging Face’s TopKLogitsWarper); 0 keeps every token.
int KeyValueHeads { get; set; } Gets or sets the number of key/value attention heads.
int LmFeedForwardDim { get; set; } Gets or sets the feature width of the language-model feed forward.
int Lookahead { get; set; } Gets or sets the lookahead.
int MaxGeneratedSpeechTokens { get; set; } The original synchronous wrapper caps a generation at 1,000 new speech tokens.
int MaxSpeechTokens { get; set; } Gets or sets the max speech tokens.
int PerceiverHeads { get; set; } Gets or sets the number of attention heads in the perceiver.
int PerceiverQueries { get; set; } Gets or sets the perceiver queries.
int RopeOriginalMaxPositions { get; set; } Gets or sets the rotary positional embedding original max positions.
int SpeechConditionPromptTokens { get; set; } Gets or sets the speech condition prompt tokens.
int SpeechStartToken { get; set; } Gets or sets the token ID for speech start.
int SpeechStopToken { get; set; } Gets or sets the token ID for speech stop.
int SpeechTokenizerHeads { get; set; } S3Tokenizer v2’s count of 64-wide rotary heads.
int SpeechTokenizerState { get; set; } S3Tokenizer v2’s state width, equal to 64 times its head count.
int SpeechVocabularySize { get; set; } Gets or sets the speech vocabulary size.
int StaticChunkSize { get; set; } Gets or sets the static chunk size.
int T3SpeakerEmbeddingDim { get; set; } Gets or sets the feature width of the t 3 speaker embedding.
int TextStartToken { get; set; } Gets or sets the token ID for text start.
int TextStopToken { get; set; } Gets or sets the token ID for text stop.
int TokenMelRatio { get; set; } Gets or sets the token mel ratio.
int VoiceEncoderHiddenDim { get; set; } Gets or sets the feature width of the voice encoder hidden.
int VoiceEncoderLayers { get; set; } Gets or sets the number of layers in the voice encoder.
int VoiceEncoderMelChannels { get; set; } Gets or sets the voice encoder mel channels.
int VoiceEncoderPartialFrames { get; set; } Gets or sets the voice encoder partial frames.
string CangjieMappingPath { get; set; } The release’s Cangjie5_TC.json, which writes Chinese characters as Cangjie codes.
string ChineseSegmenterPath { get; set; } The pkuseg model directory (spacy_ontonotes: features.msgpack, weights.npz) Chinese text is segmented with.
string HebrewDiacritizerPath { get; set; } The Dicta diacritizer directory (dicta-il/dictabert-large-char-menaked: config.json, tokenizer.json, model.safetensors) Hebrew text gets its niqqud from.
string LanguageId { get; set; } The language a multilingual model speaks (one of the release’s 23 codes: ar, da, de, el, en, es, fi, fr, he, hi, it, ja, ko, ms, nl, no, pl, pt, ru, sv, sw, tr, zh).
string RussianStresserPath { get; set; } The Russian stresser directory (russian_text_stresser’s russian_dict.db, simple_cases.pkl and its spaCy ru_core_news_sm-3.6.0 files: tokenizer, tok2vec/model, morphologizer/model, morphologizer/cfg) Russian text gets its stres…
string TokenizerPath { get; set; } Path to the release’s tokenizer.json.
string WatermarkModelPath { get; set; } The Perth model the watermark is read from: perth_net_250000.pth.tar, which ships with the resemble-perth package (perth/perth_net/pretrained/implicit/).

Methods

Method Summary
ChatterboxOptions Multilingual(ChatterboxVariant variant) Options for the 23-language release (ChatterboxMultilingualTTS): its 2,454-symbol grapheme vocabulary, otherwise the English model’s sizes.
ChatterboxOptions Turbo(bool nano) Options for Chatterbox Turbo, or Nano (ChatterboxTurboTTS.from_local’s T3Config and generate()’s defaults): GPT-2 medium (24 layers, 1,024 wide) or small (12, 768) over the 50,276 GPT-2 text tokens, 6,563 speech tokens, 375 raw prompt t…