ICosyVoiceTokenReward<T>
Interfaces · AiDotNet.TextToSpeech.CodecBased
A differentiable token-domain reward for CosyVoice 3 DiffRO (paper equations 3-7).
Known implementations: CosyVoiceToken2TextReward<T>
How It Works
The reward’s weights remain frozen during policy training, but its input derivatives must be retained. Implementations return a scalar log reward, not a detached numeric score or a waveform reward.
Properties
| Property | Summary |
|---|---|
IReadOnlyList<LayerBase<T>> Layers { get; } |
The reward’s layers, put in evaluation mode during policy optimization. |
int SpeechVocabularySize { get; } |
The number of acoustic speech-code categories accepted along the last input axis. |
Methods
| Method | Summary |
|---|---|
Tensor<T> LogReward(Tensor<T> softSpeechTokens, IReadOnlyList<int> textTokens) |
Returns the scalar log reward of soft speech tokens [frames, categories] for the target transcript. |