SteerSpeech trains a small transform for each emotion and inserts its direction into a frozen speech model. Its single preprint tests stronger expression alongside speaker identity and spoken content. English-speech experiments and anger listening studies show differing trade-offs: greater intensity does not consistently mean better preservation of the original voice.
SteerSpeech changes emotional expression by adjusting internal activations rather than retraining a whole speech generator. The method, introduced by Afsara Benazir and colleagues in a single preprint, starts with differences between neutral and emotional speech. A small learned transform refines that direction for injection into an intermediate layer. Training also protects speaker identity, transcript content and audio quality.[1]
Experiments use Qwen3-TTS-0.6B, a model that turns text into speech, with roughly thirty-three thousand transform parameters per emotion. The pipeline generates speech tokens and then replays the sequence so supervision can update only the small transform. The backbone, audio decoder and expert models remain fixed. The expert supervision and replay path are removed when generating speech for use.[1]
English-speech tests include training speakers, unseen speakers and native Mandarin speakers speaking English. The anger evaluations include two blinded listening studies involving fifty-five people. One comparison favors SteerSpeech for intensity but the simpler steering method for identity preservation. Other comparisons at stronger settings favor SteerSpeech. Happiness and sadness have different outcomes, and other languages and speech backbones remain untested.[1]
