NOESIS / AMAImedia

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: @AMAImediacom
• NOESIS version: v15.9
• Release date: 2026-08-28
• Last updated: 2026-08-29

================================================================================ NOESIS-Whisper3-1.6B-Large-Turbo-Darwin-99LANG-ONNX-INT8 -- NOESIS Bundle README

NOESIS provenance

Bundle : NOESIS-Whisper3-1.6B-Large-Turbo-Darwin-99LANG-ONNX-INT8 Parent bundle : NOESIS-3.5B-A0.5B-DUBBING-FILM BF16 source : NOESIS-Whisper-Large-V3-Turbo-Darwin-BF16 (encoder-averaged 32+32 layer merge of openai/whisper-large-v3 + openai/whisper-large-v3-turbo, v3 decoder kept, ~1.5–1.6B params) NOESIS role : INT8 ONNX deployment variant of the NOESIS encoder-averaged Whisper Darwin merge. Separate encoder + decoder ONNX graphs with external-data tensors. Multilingual ASR auxiliary — language-ID, word-level timestamping, fallback transcription across 99 Whisper-supported languages. CPU / low-VRAM onnxruntime deployable. Auxiliary to production-primary omniASR-CTC-7B (R-OMNIASR-7B-ONLY-PROD-NO-DOWNGRADE). NOESIS version : v15.9 Release date: 2026-08-28 Released : 2026-04 (BF16 source merge); 2026-05-28 (INT8 ONNX sibling) Last updated : 2026-08-29

================================================================================ Architecture

Property	Value
Architecture	Whisper encoder-decoder (Darwin encoder-averaged merge)
Encoder layers	32 (per-tensor average: v3 × 0.55 + Turbo × 0.45)
Decoder layers	32 (kept as-is from whisper-large-v3)
d_model	1280
Vocab size	51,866
Languages	99 (multilingual)
Official model class	Whisper Large-v3 / Large-v3-Turbo encoder-averaged merge
Actual source-model parameters	1,543,490,560 (~1.543B)
Deployment format	INT8 ONNX; ONNX files do not expose safetensors parameter metadata
Precision	INT8 ONNX (per-channel)
Encoder	encoder_model_int8.onnx (772 KB graph) + .data (660 MB external)
Decoder	decoder_model_int8.onnx (1.85 MB graph) + .data (1.18 GB external)
Full repository size	1.845 GB decimal (1.718 GiB)
Compression vs BF16 source	~1.6×
License	MIT (OpenAI Whisper + NOESIS attribution)
Runtime	onnxruntime (CPU / CUDA / DirectML)

Both encoder and decoder use ONNX external-data format — .data sidecars are required alongside the .onnx graphs.

================================================================================ Bundle Contents

.
├── README.md
├── LICENSE                          # MIT (OpenAI) + NOESIS notice
├── config.json                      # Whisper config
├── generation_config.json
├── preprocessor_config.json         # 128-bin log-Mel spectrogram (v3)
├── processor_config.json
├── tokenizer.json
├── tokenizer_config.json
├── special_tokens_map.json
├── added_tokens.json
├── vocab.json
├── merges.txt                       # SentencePiece BPE merges
├── encoder_model_int8.onnx          # encoder graph
├── encoder_model_int8.onnx.data     # encoder weights (~660 MB external)
├── decoder_model_int8.onnx          # decoder graph
└── decoder_model_int8.onnx.data     # decoder weights (~1.18 GB external)

================================================================================ Family

Variant	Format	Size	Role
NOESIS-Whisper-Large-V3-Turbo-Darwin-BF16	BF16	3.091 GB decimal (2.879 GiB)	encoder-averaged merge source
NOESIS-Whisper3-1.6B-...-ONNX-INT8 ← this bundle	INT8 ONNX	1.845 GB decimal (1.718 GiB)	deployment sibling

Note: a separate NOESIS-Whisper3-Large-Turbo-Darwin-99LANG-BF16 / ...-ONNX-INT8 pair (pure 4-layer-decoder Whisper-Turbo, ~809M) also lives in this parent bundle — that is a different model from this encoder-averaged 32+32-layer Darwin merge.

================================================================================ Quick Start (Optimum + onnxruntime)

from optimum.onnxruntime import ORTModelForSpeechSeq2Seq
from transformers import AutoProcessor

BUNDLE = ("B:/Downloads/Portable/NOESIS-VC-ONE/models/llm/"
          "NOESIS-3.5B-A0.5B-DUBBING-FILM/"
          "NOESIS-Whisper3-1.6B-Large-Turbo-Darwin-99LANG-ONNX-INT8")

proc  = AutoProcessor.from_pretrained(BUNDLE)
model = ORTModelForSpeechSeq2Seq.from_pretrained(BUNDLE,
                                                 provider="CPUExecutionProvider")

# audio: 16 kHz mono numpy / torch tensor
inputs = proc(audio, sampling_rate=16000, return_tensors="pt")
out = model.generate(**inputs, language="ru", task="transcribe", max_new_tokens=440)
print(proc.batch_decode(out, skip_special_tokens=True)[0])

================================================================================ Merge strategy (inherited from BF16 source)

Component	Strategy	Source
Encoder (32 layers)	Weighted average (v3 = 0.55, turbo = 0.45)	both models
Decoder (32 layers)	Kept as-is	whisper-large-v3

Turbo was distilled from v3-full → task vectors are linearly compatible. Averaging encoder combines v3's multilingual robustness with Turbo's distilled representations. Decoder from v3 preserves maximum generation quality (32 layers vs Turbo's 4). This INT8 ONNX sibling preserves the merge structure and applies per-channel INT8 quantization.

================================================================================ NOESIS context

In the NOESIS dubbing pipeline this INT8 ONNX bundle serves as a low-footprint multilingual ASR auxiliary for CPU-side language-ID, word-level timestamping, and fallback transcription. Primary production transcription remains omniASR-CTC-7B Hybrid (R-OMNIASR-7B-ONLY-PROD-NO-DOWNGRADE); this Darwin INT8 ONNX runs in parallel where CPU-only inference is required.

================================================================================ NOESIS Sealed Rules

R-MIT-OPENAI-PRESERVE Encoder-averaged derivative is redistributed under the same MIT terms as upstream OpenAI Whisper. Commercial use permitted.

R-VENDORED-INTERNAL Internal vendor copy inside parent NOESIS-3.5B-A0.5B-DUBBING-FILM bundle.

R-DUBBING-FILM-SCOPE Long-form multi-speaker film-dubbing ASR auxiliary. NOT a single-utterance benchmark target.

R-OMNIASR-7B-ONLY-PROD-NO-DOWNGRADE Whisper Darwin INT8 ONNX is a CPU-side multilingual auxiliary, not a production downgrade target. Production primary remains omniASR-CTC-7B.

R-WHISPER-FP16-AUX (extended to INT8 ONNX) Whisper bundles are auxiliary ASR; primary is omniASR-CTC-7B. This INT8 sibling extends that pattern to CPU / low-VRAM deployment.

R-NOESIS-FINAL-ARTIFACT-PATHS Canonical path: models/llm/NOESIS-3.5B-A0.5B-DUBBING-FILM/NOESIS-Whisper3-1.6B-Large-Turbo-Darwin-99LANG-ONNX-INT8/

R-NEVER-DELETE-WITHOUT-EXPLICIT-CONSENT Sealed 2026-05-21.

================================================================================ Provenance

Step	Source / output
Encoder source A	openai/whisper-large-v3 (32 layers, FP16)
Encoder source B	openai/whisper-large-v3-turbo (32 layers, distilled)
Decoder source	openai/whisper-large-v3 (32 layers, FP16)
Encoder merge	per-tensor weighted average, w(v3)=0.55, w(turbo)=0.45
Decoder strategy	dropped from Turbo, kept from v3 as-is
BF16 source bundle	NOESIS-Whisper-Large-V3-Turbo-Darwin-BF16 (3.091 GB decimal / 2.879 GiB)
INT8 ONNX export	encoder + decoder graphs with external-data tensors
Quantization	per-channel INT8
NOESIS version at BF16 merge	v14.7 (2026-04)
NOESIS version at INT8 ONNX export	v15.9 (2026-05-28)

================================================================================ Citation

@misc{radford2023whisper,
  title  = {Robust Speech Recognition via Large-Scale Weak Supervision},
  author = {Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman,
            Greg and McLeavey, Christine and Sutskever, Ilya},
  year   = {2023},
  url    = {https://openai.com/research/whisper}
}

@misc{noesis2026whisperdarwinint8,
  title  = {NOESIS DHCF-FNO :: Whisper-Large-V3-Turbo Darwin
            (encoder-averaged merge, INT8 ONNX, 1.6B)},
  author = {Bolotnikov, Ilia and AMAImedia},
  year   = {2026},
  note   = {INT8 ONNX deployment sibling of the encoder-averaged Whisper
            merge (v3 + v3-turbo, w=0.55/0.45, v3 decoder kept). MIT
            redistribution. NOESIS v15.9.},
  url    = {https://amaimedia.com}
}

================================================================================ License

MIT License. Original Whisper weights © OpenAI (2022). Encoder-averaged merge, INT8 ONNX export, packaging, NOESIS provenance: © AMAImedia (NOESIS DHCF-FNO project).

See LICENSE in this directory for full text.

================================================================================

NOESIS — Deterministic Hybrid Control Framework for Frozen Neural Operators (DHCF-FNO). Copyright (c) 2026 AMAImedia.com. All rights reserved.