VoxCPM2-KZ-Darwin-NOESIS-BF16

VoxCPM2 base TTS model with baked-in Kazakh LoRA weights (voxcpm_kaz_lora), producing a ready-to-deploy Kazakh TTS without runtime adapter loading.

Released as part of the NOESIS Professional Multilingual Dubbing Automation Platform (framework: DHCF-FNO — Deterministic Hybrid Control Framework for Frozen Neural Operators).

• Founder: Ilia Bolotnikov
• Organization: AMAImedia.com
• X (Twitter): @AMAImediacom
• LinkedIn: Ilia Bolotnikov
• Telegram: @djbionicl
• NOESIS version: v14.7
• Release date: 2026-04

---

⚠️ License notice

This model is a derivative of sozkz/VoxCPM2, released under the Apache License 2.0. The LoRA-fused derivative is distributed under the same license.

By downloading or using this model you agree to the Apache 2.0 license terms — see the LICENSE file in this repository for the full text.

---

Model summary

Property	Value
Base model	sozkz/VoxCPM2
Architecture	VoxCPM2 (LM + Encoder + DiT + AudioVAE)
LM backbone	28 layers, hidden=2048, GQA (16/2 heads), LongRoPE, vocab=73 448
Encoder	12 layers, hidden=1024, 16 heads
DiT (diffusion)	12 layers, hidden=1024, CFM euler solver
Audio VAE	16 kHz input → 48 kHz output
Format	BF16 safetensors
Merge method	LoRA fusion (W = W_base + lora_B @ lora_A)
LoRA rank	32, scale=1.0
Fused layers	160 (self_attn q/k/v/o_proj across all LM layers)
Primary language	Kazakh (KK)
Secondary language	Russian (RU)

---

LoRA fusion details

W_merged = W_base + (lora_B @ lora_A) * scale
scale = lora_alpha / lora_rank = 32 / 32 = 1.0

Fusing the LoRA eliminates adapter loading at inference time, reducing memory overhead and simplifying deployment.

---

Source models

Model	Role
sozkz/VoxCPM2	Base multilingual TTS
voxcpm_kaz_lora	Kazakh language adapter (r=32)

---

NOESIS context

In NOESIS this model serves as the Kazakh TTS teacher for knowledge distillation into TTS-10B specialist. Applied domain boost: KK×10 in soft-label weighting during KD.

NOESIS Stage	Role
Phase 1 → Stage 11 (TTS)	KK TTS teacher (KD stream A, text-side)
KD data generation	TTS teacher: w=0.25 (KK domain)

---

Provenance

Full merge trace including fused layer count in merge_provenance.json.

---

Acknowledgements & citation

Base model: VoxCPM2 by sozkz.

@misc{noesis_voxcpm2_kaz_darwin,
  title     = {VoxCPM2-KZ-Darwin-NOESIS-BF16},
  author    = {Bolotnikov, Ilia},
  year      = {2026},
  publisher = {AMAImedia},
  url       = {https://amaimedia.com}
}

@misc{noesis_v14,
  title  = {NOESIS v14.7: DHCF-FNO Multilingual Dubbing Platform},
  author = {Bolotnikov, Ilia},
  year   = {2026},
  publisher = {AMAImedia},
  url    = {https://amaimedia.com}
}