Accepted to EMNLP 2026

EmoTra-TTS Demo Page

Audio samples for the paper: "EmoTra-TTS: Smooth Intra-Utterance Emotion Transitions for Speech Synthesis", accepted to EMNLP 2026.

Paper Code

Abstract

EmoTra-TTS is an emotional TTS system that enables smooth, controllable intra-utterance emotion transitions by modeling emotions as continuous dynamics through frame-level VAD conditioning.

System Comparison

Comparing our proposed method against baseline systems. All samples are synthesized from the same text input. [Important] All texts are NOT included in the training set.

Text
Sample text 1 — The shocking text left him wide-eyed and then deeply sorrowful.
Emotion Transition
Surprised ⟶ Sad
Voice Source
♀ EmoVoice-DB-Sage
Qwen3-TTS
(a) Neutral Prompt Speech Only
MOSS-TTS
(b) Instruct Only; No Speaker Info.
EmoVoice
(c) Instruct + Neutral Prompt Speech
CosyVoice2
(d) Emo Prompt Speech + Concat
ElevenLabs v3
(h) Closed-Source Commercial TTS
gpt-4o-mini-tts
(h) Closed-Source Commercial TTS
Synthetic Data (ours)
(f) Synthetic Data
EmoTra-TTS (Proposed)
(g) Proposed EmoTra
Text
Sample text 2 — He joked with his neighbor but learning about the old dog passing crushed him.
Emotion Transition
Happy ⟶ Sad
Voice Source
♂ EmoVoice-DB-Ash
CosyVoice2
(a) Neutral Prompt Speech Only
Qwen3-TTS
(b) Instruct Only; No Speaker Info.
MOSS-TTS
(d) Emo Prompt Speech + Concat
WeSCon
(e) Emo Prompt Speech + Word-leve
ElevenLabs v3
(h) Closed-Source Commercial TTS
gpt-4o-mini-tts
(h) Closed-Source Commercial TTS
Synthetic Data (ours)
(f) Synthetic Data
EmoTra-TTS (Proposed)
(g) Proposed EmoTra
Text
Sample text 3 — The weather was so bad that day, and I was sadly walking home when I suddenly saw something that made me extremely angry.
Emotion Transition
Sad ⟶ Angry
Voice Source
♀ EmoVoice-DB-Sage
MOSS-TTS
(a) Neutral Prompt Speech Only
CosyVoice2
(c) Instruct + Neutral Prompt Speech
MOSS-TTS
(d) Emo Prompt Speech + Concat⚠ Volume Change
WeSCon
(e) Emo Prompt Speech + Word-level
ElevenLabs v3
(h) Closed-Source Commercial TTS
gpt-4o-mini-tts
(h) Closed-Source Commercial TTS
Synthetic Data (ours)
(f) Synthetic Data⚠ Volume Change
EmoTra-TTS (Proposed)
(g) Proposed EmoTra⚠ Volume Change
Text
Sample text 4 — His furious shouting stopped when a strange shadow made his blood run cold.
Emotion Transition
Angry ⟶ Fearful
Voice Source
♂ EmoVoice-DB-Ash
CosyVoice2
(a) Neutral Prompt Speech Only
MOSS-TTS
(b) Instruct Only; No Speaker Info.
EmoVoice
(c) Instruct + Neutral Prompt Speech
Qwen3-TTS
(d) Emo Prompt Speech + Concat
ElevenLabs v3
(h) Closed-Source Commercial TTS
gpt-4o-mini-tts
(h) Closed-Source Commercial TTS
Synthetic Data (ours)
(f) Synthetic Data
EmoTra-TTS (Proposed)
(g) Proposed EmoTra
Text
Sample text 5 — He cursed at the jammed printer and then chuckled when it suddenly worked again.
Emotion Transition
Angry ⟶ Happy
Voice Source
♀ EmoVoice-DB-Sage
Qwen3-TTS
(a) Neutral Prompt Speech Only
Qwen3-TTS
(b) Instruct Only; No Speaker Info.
CosyVoice2
(c) Instruct + Neutral Prompt Speech
Qwen3-TTS
(d) Emo Prompt Speech + Concat
ElevenLabs v3
(h) Closed-Source Commercial TTS
gpt-4o-mini-tts
(h) Closed-Source Commercial TTS
Synthetic Data (ours)
(f) Synthetic Data
EmoTra-TTS (Proposed)
(g) Proposed EmoTra
Text
Sample text 6 — He felt surprised at the unexpected award, then sadness when he realized a friend was overlooked.
Emotion Transition
Surprised ⟶ Sad
Voice Source
♂ EmoVoice-DB-Ash
MOSS-TTS
(b) Instruct Only; No Speaker Info.
EmoVoice
(c) Instruct + Neutral Prompt Speech
CosyVoice2
(d) Emo Prompt Speech + Concat
WeSCon
(e) Emo Prompt Speech + Word-level
ElevenLabs v3
(h) Closed-Source Commercial TTS
gpt-4o-mini-tts
(h) Closed-Source Commercial TTS
Synthetic Data (ours)
(f) Synthetic Data
EmoTra-TTS (Proposed)
(g) Proposed EmoTra
Text
Sample text 7 — She jumped at the shadow and then felt true terror as it moved closer.
Emotion Transition
Surprised ⟶ Fearful
Voice Source
♀ EmoVoice-DB-Sage
CosyVoice2
(a) Neutral Prompt Speech Only
MOSS-TTS
(b) Instruct Only; No Speaker Info.
EmoVoice
(c) Instruct + Neutral Prompt Speech
CosyVoice2
(d) Emo Prompt Speech + Concat
ElevenLabs v3
(h) Closed-Source Commercial TTS
gpt-4o-mini-tts
(h) Closed-Source Commercial TTS
Synthetic Data (ours)
(f) Synthetic Data
EmoTra-TTS (Proposed)
(g) Proposed EmoTra
Text
Sample text 8 — She smiled at the gift but reading the mocking note inside made her rage.
Emotion Transition
Happy ⟶ Angry
Voice Source
♂ EmoVoice-DB-Ash
CosyVoice2
(a) Neutral Prompt Speech Only
Qwen3-TTS
(d) Emo Prompt Speech + Concat
MOSS-TTS
(d) Emo Prompt Speech + Concat
CosyVoice2
(d) Emo Prompt Speech + Concat
ElevenLabs v3
(h) Closed-Source Commercial TTS
gpt-4o-mini-tts
(h) Closed-Source Commercial TTS
Synthetic Data (ours)
(f) Synthetic Data
EmoTra-TTS (Proposed)
(g) Proposed EmoTra

Dual-Stage SFT

Comparing with and without flow-matching SFT in the dual-stage training pipeline. Corresponds to Section 5.3 of the manuscript.

Text
Sample text — I was really disappointed about what happened, but the more I think about it, the more angry I feel.
Emotion Transition
Sad ⟶ Angry
w/o flow SFT
EmoTra ★

Fixed Scale ε

Comparing different fixed scale ε values for the direction–magnitude injection. Corresponds to Section 5.4 of the manuscript.

Text
Sample text — She smiled at the gift but reading the mocking note inside made her rage
Emotion Transition
Happy ⟶ Angry
ε = 0.04 ★
ε = 0.10
ε = 0.07 ★