cstr/phonon2-GGUF

Phonon-2 GGUF

GGUF conversions of Fermion Research's Phonon-2 for CrispASR, using its existing Parakeet TDT backend. Phonon-2 is an English speech recognizer derived from NVIDIA Parakeet TDT 0.6B v3, with retrained five-value encoder weights and 6-bit remaining tables.

File	Size (decimal MB)	Use
phonon2-f16.gguf	1,255	Reference fidelity, ordinary F16 rounding
phonon2-q8_0.gguf	674	Default balance of fidelity and size
phonon2-q4_k.gguf	402	Smaller download; more transcript drift

These are standard ggml formats. They do not preserve the upstream 164 MB transport's custom base-3/five-value packing. Upstream likewise expands weights at load time in its runtimes; download size is not runtime RAM.

Usage

With a CrispASR build containing issue #481:

crispasr -m phonon2 -f recording.wav
crispasr -m phonon2 --model-quant f16 -f recording.wav
crispasr -m phonon2 --model-quant q4_k -f recording.wav

Downloaded files also work with the existing Parakeet backend in earlier compatible CrispASR builds:

crispasr --backend parakeet -m phonon2-q8_0.gguf -f recording.wav

Python bindings auto-detect the GGUF architecture:

import soundfile as sf
from crispasr import Session

audio, rate = sf.read("recording.wav", dtype="float32")
if audio.ndim == 2:
    audio = audio.mean(axis=1)
with Session("phonon2-q8_0.gguf") as session:
    for segment in session.transcribe(audio, sample_rate=rate):
        print(segment.text)

Validation and limits

The converter's expansions matched all 723 source tensors exactly against Fermion's independent reader. F16 passed all 28 compared activation stages on JFK and the upstream LibriSpeech sample (minimum cosine 0.999996 / 0.999796; norm-ratio error bounded by 0.065% / 0.259%).

On 21 English clips through CPU C ABI sessions with an arbitrary model filename:

Export	Exact reference transcripts	Normalized word edits versus reference
F16	21/21	0/298
Q8_0	19/21	1/298
Q4_K	15/21	7/298

Reference: stock Transformers ParakeetForTDT using the unmodified upstream container reader. The clips are JFK, the upstream LibriSpeech sample, and every fourth clip (19 of 73) in hf-internal-testing/librispeech_asr_dummy's clean validation split. Normalization lowercases and ignores punctuation, retaining apostrophes inside words. These are edits against the model reference, not WER against human ground truth. Q8 changed one capitalization and one name. This small check does not reproduce the seven-set Open ASR Leaderboard results. No CrispASR Metal/CUDA/Vulkan speed or accuracy claim is made here.

Reproduce conversion

python models/convert-parakeet-to-gguf.py --hf FermionResearch/Phonon-2 --output phonon2-f16.gguf
crispasr-quantize phonon2-f16.gguf phonon2-q8_0.gguf q8_0
crispasr-quantize phonon2-f16.gguf phonon2-q4_k.gguf q4_k
python tools/dump_reference.py --backend phonon2 --model-dir FermionResearch/Phonon-2 \
  --audio samples/jfk.wav --output phonon2-jfk-ref.gguf
crispasr-diff parakeet phonon2-f16.gguf phonon2-jfk-ref.gguf samples/jfk.wav

Source revision: 9c7fef3584499a88fe8d394427f45851bbb8b446. Source archive SHA-256: 98125795b6dda72f5c6eee9ba33d19815df65dcb18b50a357bf9f73c9935309e. Expanded container SHA-256: 4b6bfa3a12cc3c4e0a54f2ab3ec4ca7a842b09e5c7ecfc8e7ca0ac6cc8c11468.

Attribution

Phonon-2: Copyright 2026 Fermion Research. Base model: NVIDIA Parakeet TDT 0.6B v3. Weights remain CC-BY-4.0. Changes here: exact unpacking of the Fermion container, tensor renaming/layout conversion to GGUF, F16 rounding, and additional ggml quantization for Q8/Q4. The original NOTICE and weights license accompany these artifacts. Fermion's container reference code is Apache-2.0 and is vendored with its license in CrispASR.