Activation Steering for Accent-Neutralized Zero-Shot Text-To-Speech

Audio samples for the paper "Activation Steering for Accent-Neutralized Zero-Shot Text-To-Speech".

We introduce a simple yet effective activation steering method for accent-neutralized zero-shot TTS. The generated speech is expected to have the same timbre as the reference speech, but with neutralized accents (US American accent is considered as the neutral accent). Experiments are conducted on Qwen3-TTS. We extract the steering vectors using the samples from L2-ARCTIC, and applied to both L2-ARCTIC and speechocean762 for evaluation.

Below we showcase the generated speech samples with and without activation steering. We show the TTS target text (Target Text), the reference speech (Reference Speech), the transcription of the reference speech (Reference Text), the generated speech without activation steering (Unsteered) and the generated speech with activation steering (Steered), which is our proposed method. The samples shown in this page are from the Qwen3-TTS 1.7B model, steered layer 15, with steering strength alpha = 1.0.

We show audio samples on two evaluation sets: L2-ARCTIC (in-domain evaluation) and speechocean762 (out-of-domain evaluation).

---

speechocean762 samples (out-of-domain evaluation)

Target Text	Reference Text	Reference Speech	Unsteered (Qwen3-TTS Baseline)	Steered (Proposed)
FIVE SEVEN ONE THREE	NOW IS THE NEW LATER
DOOR HAS ARE BEAUTIFUL SHIRTS	IT WAS ONLY A NOTE A FEW LINES
ONE FIVE NINE NINE	NOT SORRY FOR THE LANGUAGE
LOOK AT THE BIG FOOT	THIS LOOKS AS IF IT'S GOING TO WORK
MARK IS GOING TO SEE ELEPHANT	I WAS AT THE CONCERT
IT WAS A PROOF OF HIS FRIENDSHIP	BOSS WERE MADE TO BE BROKEN
MARY LIKES YOUR RED SWEATER	NO NO HE WAS TOO GOOD A WORKMAN FOR THAT
JAM LOOKED IN THE KITCHEN	THE ONLY ONE WHO KNOWS WHY IS HIM
LILLY LIVES IN CAIRO NOW	BUT STILL NO IDEA WANT WHAT TO ASK
DORA CAN SEE THE SHEEP	HE WAS STILL AT HOME

---

L2-ARCTIC samples (in-domain evaluation)

Target Text	Reference Text	Reference Speech	Unsteered (Qwen3-TTS Baseline)	Steered (Proposed)
Could the incident have anything to do with Jeanne and Pierre.	There was one difficulty however
Could the incident have anything to do with Jeanne and Pierre.	There was one difficulty however
Could the incident have anything to do with Jeanne and Pierre.	There was one difficulty however
Could the incident have anything to do with Jeanne and Pierre.	There was one difficulty however
Then you don't believe in altruism.	It's also an insidious deceitful sun
Then you don't believe in altruism.	It is also an insidious deceitful sun
Then you don't believe in altruism.	It is also an insidious deceitful sun
Then you don't believe in altruism.	It's also an insidious deceitful sun
O'Brien had been a clean living young man with ideals.	May drought destroy your crops
O'Brien had been a clean living young man with ideals.	May drought destroy your crops
O'Brien had been a clean living young man with ideals.	May drought destroy your crops
O'Brien had been a clean living young man with ideals.	May drought destroy your crops
It was the same way with our revolvers and rifles.	He was manifestly distressed by my coming
It was the same way with our revolvers and rifles.	He was manifestly distressed by my coming
It was the same way with our revolvers and rifles.	He was manifestly distressed by my coming
It was the same way with our revolvers and rifles.	He was manifestly distressed by my coming

---