Phonon-2 for Glade
Glade is a native Swift library for on-device speech inference on Apple
silicon. This repository supplies converted model assets and configuration for
Glade's model-specific host runtime. Glade's runtime source is currently private.
Three experimental conversions of Fermion Research's Phonon-2
for Apple silicon and Core AI APIs on macOS/iOS 27. LUT6 is recommended
for speed. The other variants expose the tradeoff between download size,
working memory, and inference time.
Phonon-2 is an encoder-heavy FastConformer/TDT model derived from NVIDIA
Parakeet V3. The conversions preserve its trained five-value encoder weights
and original INT6 decoder values. They use FP16 activations and the publisher's
pause-based chunking policy: 25–35 seconds, targeting 30 seconds, preserving
all audio. English is the qualified language.
| Directory |
Encoder storage |
Unpacked files |
LZRAVEN archive |
Qualification |
lut6/ |
Exact six-bit palettes shared across eight rows |
583.4 MB |
283.2 MB |
M3 Mac and iPhone 15 Pro Max |
lut4/ |
Exact four-bit palettes shared across two rows |
438.6 MB |
235.0 MB |
M3 Mac and iPhone 15 Pro Max |
size/ |
Original packed encoder records, expanded into FP16 tensor inputs at load |
217.2 MB |
178.6 MB |
M3 Mac only |
Sizes use decimal MB. Each archive restores its corresponding directory
byte-for-byte. Core AI specializes the extracted .aimodel assets on the
device; the archives contain portable source assets, not device caches.
Each directory includes its own metadata.json, runtime.json, vocabulary,
frontend constants, compact decoder data, license, and attribution.
Measured performance
JFK's “We choose to go to the Moon” speech: a 20-second excerpt and the
full 18-minute 15-second recording. Measurements were collected September 30,
2026. Larger RTFx means faster. All Core AI values are three-run medians after
preparation and warmup.
M3 MacBook Air, 16 GB, macOS 27.0.1. Includes frontend, encoder and decoding;
excludes loading, file reading and chunk planning. Full-file planning adds
about 0.06 seconds.
| Variant |
20-second excerpt |
Full speech |
Full RTFx |
Sampled full-pass memory |
| LUT6 |
0.251 s |
2.461 s |
445.0× |
1.05–1.20 GB |
| LUT4 |
0.330 s |
4.470 s |
245.0× |
0.90 GB |
| Size-focused |
0.365 s |
5.335 s |
205.3× |
2.18 GB |
Publisher fermion MLX defaults |
0.231 s |
12.668 s |
86.5× |
11.34 GB |
All memory figures use sampled process footprint plus simultaneous attributed
accelerator no-footprint memory; external or unattributed driver/compiler
memory is excluded. The publisher's 11.34 GB was measured during a dedicated
timed API run. Its unmodified CLI reached 12.09 GB across loading, warmup and
inference. Much of that was reusable allocation cache: clearing it after
completion reduced the CLI footprint to 2.30 GB with active tensors unchanged.
No lower-cache performance claim is implied.
The unmodified publisher runtime was fermion-research 0.2.4 with
tdt16,dense16. Its timing includes chunk planning. A separate invocation of
its own benchmark CLI confirmed a 12.537-second full-speech median (87.4×).
The short excerpt favors the publisher runtime; these speedups describe the
long-recording workload.
iPhone 15 Pro Max, A17 Pro, iOS 27.0.1, Release build. Includes bounded WAV
reading, chunk planning, frontend, encoder and decoding; preparation and report
serialization are excluded. Foreground, nominal thermals, Low Power Mode off.
| Variant |
20-second excerpt |
Full speech |
Full RTFx |
Sampled full-pass memory |
| LUT6 |
0.321 s |
3.075 s |
356.2× |
1.03 GB |
| LUT4 |
0.370 s |
5.349 s |
204.8× |
0.90 GB |
| Size-focused |
— |
— |
— |
Not measured on iPhone |
Preparation and placement
| Variant |
Observed first preparation, M3 |
Cached preparation, M3 |
Observed first preparation, iPhone |
| LUT6 |
104.2 s |
0.35 s |
72.6 s |
| LUT4 |
179.1 s |
0.18–0.66 s |
243.1 s |
| Size-focused |
49.7 s |
3.56–3.59 s |
Not measured |
These are observed first-use sessions, not repeated clean-install medians.
The LUT4 Mac run reused decoder caches. LUT6 and size-focused runs reused
unchanged subsampling and decoder specializations. The size-focused variant
expands about 1.21 GB of FP16 encoder tensors at load, accounting for most of
its cached preparation time and extra working memory. Its initial preparation
was measured before a benchmark-reporting issue was corrected; subsequent
clean inference runs established its numerical and performance results.
Complete warmed traces for each qualified platform/variant show 493 ANE
predictions and zero target-process GPU intervals: 259 predictions during
subsampling/encoding and 234 during batched decoding. CPU feature extraction
and host TDT control remain outside these graphs. Placement is evidence of
ANE execution, not a measurement of arithmetic utilization.
Accuracy and runtime contract
All three conversions produce identical token IDs and transcript text on the
two reference clips. The phone-qualified variants match their Mac outputs.
Full-speech WER is 2.93% using Whisper's English normalizer (65 errors in
2,220 reference words), equal to an independent FP32 reconstruction's normalized
word sequence; the publisher MLX defaults score 3.02%. This is one recording,
not a general accuracy ranking. Two full-speech chunks differ in tokenization
from FP32 while preserving the normalized words.
Native word timings are returned. They are finite, ordered and within each
chunk; human word-boundary accuracy has not been measured independently. The
size-focused variant shifts eight full-speech word intervals relative to the
palette variants, by at most 160 ms, without changing the transcript.
Use a runtime that implements the included metadata and scheduling policy.
The six encoder graphs each evaluate four layers over the complete
448-position context. Valid-length masks, relative-position conventions,
silence handling, TDT state and duration decoding, and compact weight expansion
are part of the runtime contract. runtime.json specifies four concurrent
encoder requests and ANE decoding of up to 128 independent audio chunks.
The size-focused variant uses a single batched decoder asset for both single
utterances and batches; its encoder weights are explicit tensor inputs.
Download
Download one variant. On macOS 27, the smaller archive can be extracted with
Apple's aa tool:
hf download coder543/phonon-2-glade-experimental archives/phonon-2-lut6.aar --local-dir phonon-download
mkdir -p phonon-2-lut6
aa extract -i phonon-download/archives/phonon-2-lut6.aar -d phonon-2-lut6
For individual runtime files:
hf download coder543/phonon-2-glade-experimental --include 'lut6/*' --local-dir phonon-download
Choose lut4 or size in those paths for another design. Applications using
compressed delivery need an AppleArchive/LZRAVEN extraction step before Core AI
loading. Hugging Face supplies file checksums through its snapshot metadata.
Attribution
Weights are CC-BY-4.0, derived from NVIDIA Parakeet TDT 0.6B V3 and
retrained/quantized by Fermion Research. The original license and NOTICE are
included. Conversion changes include staged Core AI graphs, FP16 activations,
exact palette or external-weight storage, compact decoder packing, and batched
decoding. No retraining or pruning is performed. Compiler source-location
metadata is removed from the published assets.
Source revision: 9c7fef3584499a88fe8d394427f45851bbb8b446.
Tokenizer/configuration revision: 541d1f99c6b0c3cd0b11a95167540bb8edefd82b.
These identify conversion provenance; consumers should resolve a consistent
current repository snapshot and honor its runtime metadata.
