●builderYou can use SAE latent steering as a lightweight control mechanism for TTS attributes like laughter and speech rate without retraining the base model.
●researcherThis extends SAE interpretability methods to multimodal token streams, with a modality-aware labeling pipeline that could generalize to other text-audio LMs.
●designerTargeted latent interventions give you fine-grained expressive control over TTS output — laughter, gender, rate — without prompt engineering or model swaps.