Abstract
Speech becomes less intelligible in noisy environments, and humans naturally adapt their voice to compensate. Inspired by this behavior, we investigate whether a text-to-speech (TTS) model can be guided to produce more intelligible speech using activation steering, without retraining. We focus on two characteristics of the Lombard effect: increased vocal effort and hyper-articulation. We introduce a prompt-relative steering mechanism that prevents steering effects from accumulating during generation while allowing their strength to be adjusted dynamically. Across seen and unseen speakers and multiple languages, our method produces systematic changes in Lombard-related acoustic features, preserves speaker similarity (90–95%), and reduces WER under background noise by 7–22% at 1 dB SNR. These results show that pretrained TTS models can be dynamically controlled to generate more intelligible speech without retraining.
Audio Samples
Expresso Dataset
Comparison across different steering strengths and configurations.
| Condition | Sample #1 | Sample #2 | Sample #3 | Sample #4 |
|---|---|---|---|---|
| Regular (Baseline) | ||||
| Steering (λ = 0.5) |
||||
| Steering (λ = 1.0) |
||||
| Steering (λ = 2.0) |
||||
| Vocal Effort Only | ||||
| Articulation Only |
Speech in Noise Evaluation
Comparison between unsteered baseline (left) and our steered method (right) under background noise.
Multilingual Evaluation
Cross-lingual transfer performance (Regular vs. Steered).
German Evaluation
Spanish Evaluation
Japanese Evaluation
Dynamic Activation Steering
Demonstrating temporal control with dynamic steering windows applied during streaming generation.