UniverSR-Darwin-NOESIS-BF16

Weighted-average merge of UniverSR-speech and UniverSR-audio, producing a generalist 48kHz audio super-resolution model optimized for speech dubbing. Converted from FP32 (pytorch_model.bin) to BF16 safetensors format.

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

---

Model summary

Property	Value
Architecture	UniverSR (ConvNeXt flow-matching, CFM)
Parameters	~57.2M
Original format	FP32 (pytorch_model.bin, 229 MB)
Stored format	BF16 safetensors (model.safetensors, 115 MB)
Sample rate	48 kHz
Merge method	Weighted average
Primary domain	Speech (dubbing output)
Secondary domain	General audio

---

Source models

Model	Role	Weight
UniverSR-speech	Speech-optimized SR	0.60
UniverSR-audio	Universal audio SR	0.40

Higher weight on speech variant aligns with NOESIS primary use case: professional dubbing output enhancement.

---

NOESIS context

Used in NOESIS Phase 1 (11-18) pipeline as the audio super-resolution post-processing step before final DSP assembly. Upsamples synthesized speech to 48 kHz professional quality.

---

Provenance

Full merge trace in merge_provenance.json.

---

Citation

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