Phonon-2 Inference Model Pack (ipsilondev/Phonom2-fermion)

Stand-alone inference package and Git submodule containing weights, manifests, and offline tokenizer for Phonon-2 (2.1-bit ternary FastConformer 0.6B + TDT ASR).

Files Included

File	Size	Purpose
model.fermion	169.2 MB	2.1-bit quint5 packed model container (24-layer FastConformer + 2-layer LSTM decoder + joint)
config.json	277 KB	Architecture hyperparameters, mel filterbank settings, and model metadata
packed_manifest.json	821 B	Manifest indexing packed ternary tensor offsets, trit planes, and scale vectors
bps_manifest.json	799 B	Fermion package manifest
tokenizer.json	1.1 MB	Offline vocabulary and SentencePiece BPE tokenizer (8,193 tokens)
tokenizer_config.json	352 B	Tokenizer configuration
processor_config.json	408 B	Audio preprocessor configuration (128 mel bins, 16kHz)
reference_transformers.py	6.8 KB	Stand-alone PyTorch / Transformers model loader
fermion_container.py	3.9 KB	Stand-alone container decompressor and trit dequantizer

---

Submodule Usage

Add this repository as a Git submodule in your downstream projects:

Via Hugging Face (Recommended for large model weights)

git submodule add https://huggingface.co/ipsilondev/Phonom2-fermion models/phonon2
git submodule update --init --recursive

Via GitHub

git submodule add https://github.com/ipsilondev/Phonom2-fermion.git models/phonon2
git submodule update --init --recursive
git lfs pull

---

Python Inference Quickstart

import torch
import reference_transformers as rt

# Load model using local container and tokenizer (completely offline)
model, processor, _ = rt.load_model(
    container_path="models/phonon2/model.fermion",
    base_dir="models/phonon2",
    dtype=torch.float32
)
model.eval()

# Process audio
import soundfile as sf
audio, sr = sf.read("test.wav")
inputs = processor([audio], sampling_rate=16000, return_tensors="pt")

with torch.no_grad():
    out = model.generate(input_features=inputs["input_features"])

text = processor.batch_decode(out.sequences, skip_special_tokens=True)[0]
print("Transcript:", text)

License

Model weights: CC-BY-4.0 (derived from NVIDIA Parakeet TDT 0.6B). Code & tools: Apache-2.0.