Emotion sweep — soft instruct control

Same passage, same engines, only the instruct= prompt changes. Pattern: [emotion]な大人の朗読 — keeps the "adult narrator" frame constant so the emotion descriptor is the only varying axis.

Test passage:
最近、新しい本を読み始めました。物語は静かな町から始まり、主人公はゆっくりと自分の過去と向き合っていきます。読み終える頃には、季節も変わっているかもしれません。

"Recently I started reading a new book. The story begins in a quiet town, and the protagonist slowly comes to terms with their own past. By the time I finish reading, the seasons may have changed."
Emotion / instruct prompt
Qwen3-TTS (Ono_Anna · MLX 8-bit)
VoxCPM2 (default voice · MLX 4-bit)
Calm 穏やかな大人の朗読 "calm adult narration"
Sad 悲しげな大人の朗読 "sad adult narration"
Happy 嬉しそうな大人の朗読 "joyful adult narration"
Angry 怒りを込めた大人の朗読 "angry / with-anger adult narration"
Anxious 不安げな大人の朗読 "anxious adult narration"
What to listen for: Whole-utterance prosodic shift, not per-word emphasis. Soft instruct mode = the model's interpretation of the descriptor — there's no intensity dial. If an emotion sounds flat or undifferentiated, that's the limit of natural-language conditioning vs explicit emotion tokens.

The next level up is IndexTTS-2 (Bilibili, late 2025) which has disentangled timbre + emotion with quantitative intensity — separate clip drives the emotion. Tradeoffs: non-Apache license (commercial use needs Bilibili sign-off), optimized for English not JP, ~15× realtime RTF on M-series Mac.