StepAudioTrainingStage

Enums · AiDotNet.TextToSpeech.MultiModal

Step-Audio’s training stages (§4.2 and §5).

Fields

Field Summary
StepAudioTrainingStage ChatFineTuning AQTA supervised fine-tuning (§5.2.2): audio question, text answer, cosine 5.656e-5 → 5.656e-6.
StepAudioTrainingStage Ppo PPO against the reward model (§5.2.6): clip 0.2, KL 0.05, the critic warmed up 80 steps, cosine 1e-6 → 2e-7.
StepAudioTrainingStage PreTraining1 Pre-training stage 1: audio continuation (and text), the 5,120 new audio tokens’ embedding and head rows at 5× the backbone’s 2e-5.
StepAudioTrainingStage PreTraining2 Stage 2: audio continuation and audio-text interleaved data, the same rates.
StepAudioTrainingStage PreTraining3 Stage 3: continuation, interleaving, ASR and TTS data; every rate on one cosine 2e-5 → 5e-6.
StepAudioTrainingStage RewardModel The reward model (§5.2.4): Bradley–Terry on preference pairs, cosine 1.24e-5 → 6e-6.
StepAudioTrainingStage TtsFineTuning The TTS model’s supervised fine-tuning (§5.1.2): the two-turn chat format, one epoch, cosine 2e-5 → 2e-6.
int value__