Emotion LoRAs for MOSS-TTS Voice-Acting (4.55B)

One LoRA adapter per EmoNet emotion, each the best of a rank {16,32,64} sweep on natural (EmoLia) + intense Got-Talent speech, with ~25% same-voice references and procedural emotion captions. Click an emotion to hear five neutral sentences in English / German / French, with and without an emotion reference. z = EmoNet emotion intensity of the output (baseline = v2 with no LoRA); blend = vocal-burst-blend; genu = genuineness; spk = speaker similarity.

Adapters on Hugging Face: TTS-AGI/moss-emotion-loras-40.

Merge experiments: ▶ LoRA merge-intensity scaling · ▶ combining two emotion LoRAs · ▶ diff-vectors: per-dimension max/min LoRA

Character-voice LoRAs: ▶ Race voice LoRAs — 20 game-character races (rank 64) — one adapter per race trained on chavo character voices; two neutral sentences with no reference, so you hear the pure race timbre. Adapters: TTS-AGI/moss-race-loras.

Sidon enhancement: ▶ Character voices — before / after Sidon — every character demo re-processed with Sidon speech restoration (w2v-BERT LoRA → DAC, 48 kHz), original vs enhanced side by side with DNS-MOS scores.

VoiceNet dimension LoRAs: ▶ VoiceNet dimensions — HIGH vs LOW (throat, cartoon, metallic, nasal, chest, warmth, whisper, authority …) — the same neutral sentence generated with a high-end and a low-end adapter per perceptual dimension, so you hear each dimension move in isolation.