LAION · MOSS-8B voice-acting · prompting ablation

How to prompt MOSS-8B into extreme emotion

What this is (for dummies). We asked: which wording of the instruction makes the voice-acting model actually sound furious / terrified / heartbroken / ecstatic / disgusted / astonished? We generated 4 takes for every combination of 6 emotions × 20 instruction styles × 3 emotional texts (1,440 clips, no reference audio), and scored every clip with Empathic-Insight-Voice-Plus (emotion-annotations): the target emotion score, arousal (energy), valence, expressiveness, plus the average volume (rms in dB) of the clip. two extra arms test which instructions maximise speaker similarity to a reference voice and genuineness.

Reading the tables: higher = more of that thing. "target" = the Empathic-Insight score of the emotion we asked for (e.g. Anger head for anger prompts). "ctrl" rows are the neutral-instruction control. rms dB is negative — closer to 0 = louder.

1 · Per emotion: which instruction style works best?

anger

stylewhat it doestargetarousalrms dBexpressiveinvWER
s15_contrast ★negate calm ('not calm, not measured')3.3792.951-20.82.5240.96
s04_extreme_adjextreme adjectives, 'pushed to the limit'3.3112.960-21.22.4820.98
s01_minimaljust the emotion word3.2982.946-20.42.4450.98
s02_plainplain 'a person speaking, X'3.2753.102-22.82.4590.97
s17_arousal_wordsadrenaline/arousal vocabulary3.2282.884-22.32.4290.91
s19_kitchen_sinkeverything combined3.1882.779-22.42.3440.98
s07_shoutingexplicit shouting/screaming cues3.1842.987-21.82.4010.95
s12_imperativesecond-person commands: 'be furious!'3.1622.855-23.42.5470.92
s20_neutral_ctrl (control)neutral control3.0922.803-22.42.5000.91
s15_contrast — best take
target4.249arousal3.388rms-22.5dB
s04_extreme_adj — best take
target4.151arousal3.387rms-20.1dB
s20_neutral_ctrl — best take
target3.952arousal3.042rms-22.5dB

fear

stylewhat it doestargetarousalrms dBexpressiveinvWER
s11_escalation ★instruct an escalation arc2.8373.057-22.52.2560.97
s13_ellipsesellipses + exclamation marks2.7923.114-22.22.2570.90
s07_shoutingexplicit shouting/screaming cues2.6813.036-21.52.4690.91
s06_physicalphysical body cues2.6002.917-20.72.5100.95
s19_kitchen_sinkeverything combined2.5932.922-21.32.1710.90
s10_actor_coachmethod-actor coaching language2.4952.825-24.42.1820.95
s16_movie_climaxoscar-movie climax framing2.4472.822-21.72.3290.93
s15_contrastnegate calm ('not calm, not measured')2.4062.925-20.42.4930.91
s20_neutral_ctrl (control)neutral control1.9502.694-21.72.5700.89
s11_escalation — best take
target3.400arousal3.160rms-20.6dB
s13_ellipses — best take
target4.425arousal3.946rms-24.0dB
s20_neutral_ctrl — best take
target3.920arousal3.586rms-20.5dB

sadness

stylewhat it doestargetarousalrms dBexpressiveinvWER
s13_ellipses ★ellipses + exclamation marks2.3621.282-23.82.1910.98
s19_kitchen_sinkeverything combined2.2081.394-22.71.9150.97
s04_extreme_adjextreme adjectives, 'pushed to the limit'2.1171.285-23.92.3640.99
s10_actor_coachmethod-actor coaching language2.0401.314-24.62.0510.97
s06_physicalphysical body cues2.0331.269-23.32.1500.94
s11_escalationinstruct an escalation arc2.0241.391-23.52.2380.99
s05_repetitionrepeat the emotion word many times2.0171.351-22.12.2670.94
s12_imperativesecond-person commands: 'be furious!'2.0061.278-24.62.1620.97
s20_neutral_ctrl (control)neutral control1.8671.099-24.62.2070.94
s13_ellipses — best take
target3.205arousal1.263rms-25.1dB
s19_kitchen_sink — best take
target3.380arousal1.578rms-20.6dB
s20_neutral_ctrl — best take
target3.012arousal0.898rms-20.2dB

joy

stylewhat it doestargetarousalrms dBexpressiveinvWER
s13_ellipses ★ellipses + exclamation marks3.2893.085-23.32.5520.97
s07_shoutingexplicit shouting/screaming cues3.2323.063-21.22.8790.95
s10_actor_coachmethod-actor coaching language3.2213.199-21.52.9470.98
s19_kitchen_sinkeverything combined3.1783.190-20.82.6380.98
s05_repetitionrepeat the emotion word many times3.1133.062-22.72.5810.94
s09_scenarioground it in a concrete scenario3.0963.029-23.62.7690.99
s15_contrastnegate calm ('not calm, not measured')3.0793.079-21.22.8080.98
s12_imperativesecond-person commands: 'be furious!'3.0673.038-22.32.7430.88
s20_neutral_ctrl (control)neutral control3.0473.084-22.02.8200.97
s13_ellipses — best take
target5.110arousal4.309rms-28.9dB
s07_shouting — best take
target5.260arousal4.298rms-21.2dB
s20_neutral_ctrl — best take
target4.867arousal4.305rms-21.9dB

disgust

stylewhat it doestargetarousalrms dBexpressiveinvWER
s08_parenthetical ★stage directions in (parentheses)1.4092.526-22.72.1520.98
s20_neutral_ctrlneutral control (no emotion)1.3782.179-22.82.1630.98
s14_loudnessexplicit maximum-volume wording1.3502.512-20.12.0930.96
s11_escalationinstruct an escalation arc1.2652.403-21.12.3050.94
s06_physicalphysical body cues1.2602.314-22.02.2280.90
s19_kitchen_sinkeverything combined1.2112.307-23.12.1680.90
s13_ellipsesellipses + exclamation marks1.1492.207-23.12.3250.87
s04_extreme_adjextreme adjectives, 'pushed to the limit'1.1322.224-23.52.2630.94
s20_neutral_ctrl (control)neutral control1.3782.179-22.82.1630.98
s08_parenthetical — best take
target2.065arousal2.288rms-33.4dB
s20_neutral_ctrl — best take
target2.820arousal3.004rms-20.9dB
s20_neutral_ctrl — best take
target2.820arousal3.004rms-20.9dB

surprise

stylewhat it doestargetarousalrms dBexpressiveinvWER
s20_neutral_ctrl ★neutral control (no emotion)2.3122.660-20.62.2590.98
s13_ellipsesellipses + exclamation marks2.1742.701-21.42.4490.97
s10_actor_coachmethod-actor coaching language2.0872.672-20.52.4450.91
s16_movie_climaxoscar-movie climax framing2.0702.526-20.42.5570.92
s06_physicalphysical body cues2.0202.528-22.42.3170.93
s18_vocal_descdescribe the raw unstable voice1.9942.478-20.62.3840.99
s14_loudnessexplicit maximum-volume wording1.9552.462-21.82.5630.95
s03_dramabox_basedramabox baseline sentence1.9522.616-20.62.4790.89
s20_neutral_ctrl (control)neutral control2.3122.660-20.62.2590.98
s20_neutral_ctrl — best take
target4.375arousal2.890rms-17.1dB
s13_ellipses — best take
target3.329arousal2.417rms-21.6dB
s20_neutral_ctrl — best take
target4.375arousal2.890rms-17.1dB

2 · Overall expressiveness leaderboard (all 6 emotions pooled)

expr_index = mean z-score of target emotion + arousal + volume + expressiveness. What reliably cranks the model up — and what does nothing.

stylewhat it doesexpr indextargetarousalrms dBinvWER
s11_escalationinstruct an escalation arc0.862.3362.577-22.50.96
s14_loudnessexplicit maximum-volume wording0.592.2222.513-21.90.93
s16_movie_climaxoscar-movie climax framing0.542.2742.462-22.10.95
s19_kitchen_sinkeverything combined0.512.3832.532-21.80.95
s04_extreme_adjextreme adjectives, 'pushed to the limit'0.462.2832.507-22.30.96
s07_shoutingexplicit shouting/screaming cues0.442.2532.509-21.90.95
s13_ellipsesellipses + exclamation marks0.352.4092.521-22.70.95
s02_plainplain 'a person speaking, X'0.332.1832.474-22.30.93
s06_physicalphysical body cues0.312.2912.467-21.90.93
s01_minimaljust the emotion word0.182.2162.492-22.20.94
s15_contrastnegate calm ('not calm, not measured')0.132.2952.470-22.60.93
s03_dramabox_basedramabox baseline sentence0.062.0972.521-22.40.93
s10_actor_coachmethod-actor coaching language-0.202.2642.524-22.80.95
s12_imperativesecond-person commands: 'be furious!'-0.612.2192.451-23.10.91
s09_scenarioground it in a concrete scenario-0.612.1522.505-23.20.96
s18_vocal_descdescribe the raw unstable voice-0.652.1572.386-22.30.96
s17_arousal_wordsadrenaline/arousal vocabulary-0.692.1842.406-22.60.92
s08_parentheticalstage directions in (parentheses)-0.802.1042.430-21.90.87
s05_repetitionrepeat the emotion word many times-1.192.1572.387-23.00.91

volume ↔ arousal correlation r = 0.26; volume ↔ target emotion r = 0.09. loudest style: s19_kitchen_sink, quietest: s09_scenario.

3 · Speaker similarity: how to prompt "talk exactly like the reference"

templatespk similarity (mean)best takeinvWERgenu
p1_empty0.9310.9680.861.65
p7_neutral_ctrl0.9300.9740.881.88
p2_exact0.9290.9660.841.77
p4_same_person0.9280.9610.631.52
p5_identical0.9270.9700.822.10
p8_do_not_change0.9240.9630.911.99
p3_clone0.9200.9540.731.81
p6_session0.9070.9590.831.69

similarity = cosine of Chatterbox VoiceEncoder embeddings vs the reference. best template: p1_empty. per reference voice: chris 0.883 · goblin 0.939 · samantha 0.952

p1_empty (samantha) — best take
spk sim0.968
p1_empty (samantha) — best take
spk sim0.968
p7_neutral_ctrl (samantha) — best take
spk sim0.974

4 · Genuineness: which template sounds most real?

templategenuineness (VoiceCLAP)authenticity (EI-Plus)blendinvWER
g1_control1.9563.2561.960.94
g2_genuine1.6413.2881.670.93
g6_anti_acting1.5513.2272.130.97
g8_voice_memo1.5393.4042.840.99
g3_hesitations1.5383.3612.580.97
g4_documentary1.4903.3192.600.99
g7_method1.3963.3242.890.93
g5_spontaneous1.3383.3422.920.99

two independent judges: the VoiceCLAP genuineness head and the Empathic-Insight Authenticity head (agreement r = -0.03). best template: g1_control.

g1_control — best take
genu4.338auth3.252
g1_control — best take
genu4.338auth3.252