best-vox (research)

Swamiji's voice in English, Telugu and Kannada from one VoxCPM2 model, no reference clip. The VoxCPM2 pick since 2026-11-02 (with sw-voice/best-2-multilingual-maya); first named voxcpm2-stageB-te-reuse87h-navaratri-te-kn-baselm3e-4 (experiment 06/27, 87 h stage B). Stage A trained on 87 h of Telugu talks and shlokas (sw-voice/swamiji-telugu-reuse-train, RE-USE) plus ~3 h of English twice; this stage B adds Navaratri te/kn. Voice similarity (WavLM-ECAPA, % of real-vs-real): te 96.6, kn 97.2, en 96.8; Telugu Q&A 98.6. Replaces sw-voice/voxcpm2-stage3-te-kn-reuse-tmu1.0 (was best-2-multilingual-vox).

Full fine-tune of sw-voice/voxcpm2-stageA-te-reuse87h-en2x-baselm3e-4-acc16-1ep with the official VoxCPM trainer. Data: sw-voice/swamiji-navaratri-telugu-reuse-themes + -kannada-reuse-themes (127 + 124 whole-theme units, RE-USE), 5 epochs, batch 1 x 8, nothing frozen, flow-matching t = sigmoid(N(1.0, 1)), LR 5e-5 with base_lm.* at 3e-4. Micro-batch 1 x accumulation 8, 5 epochs = 160 steps, LR 5e-5 (base_lm.* 3e-4) cosine, warmup 24, grad clip 1.0, all parameters trained. Reference-free: the voice is in the weights. Code: swamiji-tts research/exp-iterations/06-voxcpm2-swamiji-stages.