Our EmoSteer-TTS approach enables training-free and fine-grained emotion control through activation steering:

Figure 1: Overview of the proposed EmoSteer-TTS. Emotion steering vectors and steering weights are derived from pairs of neutral and emotional reference speech samples. During inference, these vectors are used to modulate the activations of the TTS model, effectively guiding it to synthesize speech that reflects the desired emotion.

Emotion conversion supports six basic emotions: anger, disgust, fear, happiness, sadness, and surprise.

Target Emotion: Anger

Target Emotion: Disgust

Target Emotion: Fear

Target Emotion: Happiness

Target Emotion: Sadness

Target Emotion: Surprise

EmoSteer-TTS can also interpolate between a neutral reference and a target emotion to create a smooth transition:

Target Emotion: Anger

Target Emotion: Disgust

Target Emotion: Fear

Target Emotion: Happiness

Target Emotion: Sadness

Target Emotion: Surprise

The Emotion Interpolation Ability of Training-Based Baseline Methods

All the samples are grabbed from their official demo pages.

EmoSphere++ (Cho et al. 2025) offers intensity control over four emotions, i.e., Anger, Happiness, Sadness, and Surprise.

HED-TTS (Inoue et al. 2025) offers intensity control over four emotions, i.e., Anger, Happiness, Sadness, and Surprise.

EmoDubber (Cong et al. 2025) offers intensity control over three emotions, i.e., Anger, Sadness, and Surprise.

EmoSteer-TTS can also erase the emotion of a reference speech to create a neutral speech:

Emotion to Erase: Anger

Emotion to Erase: Disgust

Emotion to Erase: Fear

Emotion to Erase: Happiness

Emotion to Erase: Sadness

Emotion to Erase: Surprise

EmoSteer-TTS enables simultaneously adding two emotions to the synthesized speech. All the examples below are generated using the F5-TTS backbone.

Note: Oversteered (e.g., $\alpha$ > 2.0) samples may exhibit artifacts, noise, pitch and timbre change, or unnaturalness.

Emotions to Add: Anger and Sadness

Emotions to Add: Anger and Disgust

Emotions to Add: Fear and Surprise

Note: You may need to adjust the volume levels for better listening experience.

Emotion Interpolation

In this subsection, we present additional visualizations of F0 contours to illustrate the fine-grained and continuous emotion interpolation capabilities of the proposed EmoSteer-TTS. As shown in Figure 2, voices with angrier, happier, or more surprised tones tend to exhibit higher pitch, while sadder tones are generally associated with lower pitch. In contrast, pitch variations in disgust and fear interpolation show no clear monotonic trend, which we attribute to the fact that these emotions are more closely tied to semantic content than to acoustic characteristics.

Figure 2: Visualizations of F0 contours in emotion interpolation. From left to right: F5-TTS, E2-TTS, and CosyVoice2. From top to bottom: anger, disgust, fear, happiness, sadness, and surprise. All the synthesized speech samples are interpolated between neutrality ($\alpha$=0) and a target emotion ($\alpha$=2).

Emotion Erasure

In this subsection, we present additional F0 contour visualizations to demonstrate the emotion erasure capability of the proposed EmoSteer-TTS. As shown in Figure 3, the pitch contours of angry, disgusted, happy, and surprised voices become noticeably flatter after emotion erasure, indicating a calmer prosodic pattern. In contrast, the changes in pitch for fear and sadness are more diverse and less predictable. This variability may stem from the fact that fear can be expressed through multiple vocal styles, such as a low, trembling voice or a high-pitched scream, making it difficult for pitch alone to capture the underlying emotional shift. Similarly, sadness may manifest as either soft weeping or loud crying, resulting in inconsistent pitch patterns that do not reliably reflect emotional intensity.

Figure 3: Visualizations of F0 contours in emotion erasure. From left to right: F5-TTS, E2-TTS, and CosyVoice2. From top to bottom: anger, disgust, fear, happiness, sadness, and surprise. All the synthesized speech samples are emotionally erased from a target emotion ($\beta$=0) towards neutrality ($\beta$=2.5).

We presented EmoSteer-TTS, the first training-free framework for fine-grained, continuous, and interpretable emotion control in speech synthesis. By steering a subset of internal activations in a TTS model, our method enables flexible emotional manipulation, including emotion conversion, interpolation, and erasure, without modifying or fine-tuning the pretrained TTS model. We also constructed a curated emotional speech dataset to support steering vector construction. Extensive experiments confirm that EmoSteer-TTS achieves robust, zero-shot emotion control with broad applicability, outperforming SOTA methods. The analysis also offers deeper insights into the emotion steering dynamics of flow matching-based TTS. To the best of our knowledge, this is the first fine-grained EC-TTS approach that can transform previously uncontrollable TTS models into emotionally controllable ones without any retraining, fine-tuning, or model redesign.