MoshiTrainingStage

Enums · AiDotNet.TextToSpeech.MultiModal

Moshi’s training stages (§4.4, Table 1).

Fields

Field Summary
MoshiTrainingStage FisherFineTuning Fine-tuning on Fisher’s real two-speaker conversations.
MoshiTrainingStage InstructFineTuning Instruct fine-tuning on synthetic dialogues, the user’s stream augmented.
MoshiTrainingStage PostTraining Post-training on diarized speech split into two streams, text delay 0, text-only batches 10% of the time.
MoshiTrainingStage PreTraining Pre-training on unsupervised single-stream speech (text masked 30%, text delay ±0.6 s), half the batches text-only with their own optimizer state.
int value__