Two new research papers introduce "activation steering" as a method to control aspects of generated speech without retraining existing text-to-speech (TTS) models. The first paper, "SteerSpeech," focuses on controlling emotional expression by injecting steering vectors into model activations, demonstrating improved emotion control and speaker identity preservation with models like Qwen3-TTS. The second paper, "Loud and Clear," applies a similar prompt-relative steering mechanism to enhance speech intelligibility in noisy environments, inspired by the human Lombard effect, and shows a reduction in word error rate under background noise. AI
IMPACT Activation steering offers a promising avenue for fine-tuning TTS models for specific emotional expressions and improved intelligibility in challenging acoustic conditions without costly retraining.
RANK_REASON Two academic papers published on arXiv introducing novel methods for controlling TTS models.
- arXiv
- Dynamic Activation Steering
- Lombard effect
- Loud and Clear
- Text-to-speech (TTS)
- Hugging Face
- Qwen3-TTS
- SteerSpeech
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →