coder543/underdog-bark-3.0-plus-glade

Underdog Bark 3.0 Plus for Glade

Converted model assets and runtime configuration for Glade.

Conversion of Conway Research's Underdog Bark 3.0 Plus, source revision d4fab2f8b95681e5ad3cfb4e8c5f4e0c009b3deb. The weights retain CC-BY-4.0; see LICENSE and NOTICE. This conversion is not endorsed by Conway Research, Moondream or NVIDIA.

An English FastConformer/TDT model descended from Parakeet Ultra: 24 encoder layers, 1,024 channels, eight attention heads, and a two-layer 640-channel LSTM predictor. Exact parameter count: 627,008,134. It publishes no VAD head. Plus and Pro are separate checkpoints of the same architecture, with different weight representations and reported accuracy.

Bundle

One self-contained 591.1 MB bundle, compared with the source checkpoint's 179.5 MB download. These are converted Core AI graphs, not the upstream storage container. Requires Glade on physical Apple silicon with iOS 27 or macOS 27. Graphs consume features and recurrent states; the host handles raw audio, chunk planning, greedy TDT decoding, tokenization and native word timing.

Original five/seven-value encoder row palettes (two/three learned magnitudes, their negatives and zero) are stored losslessly in grouped LUT6 tables (eight source rows per group). Native INT5/6 codes and original scales remain in the compact host-decoder sidecar; no new centroids are fitted.

• Mono 16 kHz input, 128-bin log-mel frontend, and the source's 250 ms leading silence per independent chunk. Returned timestamps remove this padding.
• One 768-position encoder shape, split into six four-layer assets. Full bidirectional attention and the complete finite relative-position basis remain available within each chunk; staging does not split the attention context.
• Contiguous quiet-boundary chunks up to 60 seconds retain every audio sample, including silence. This bounds long-recording memory; it differs from passing an entire long recording through the source in one context.
• Two concurrent encoder requests; a batched FP16 ANE decoder supports capacities 16 and 32. Greedy decoding follows upstream defaults, with durations 0–4 and at most ten symbols per acoustic frame.
• SiLU uses equivalent exp/division math. Convolution masks precede the bias-free GLU projection. Graph rewrites pass an FP32 source check before native testing.
• metadata.json describes numerical conventions and shapes; runtime.json specifies scheduling. Native TDT word timings have an 80 ms frame step and are not forced alignment.

Mac qualification

M3 MacBook Air, 16 GB; macOS 27.0.1 (26A434); October 24, 2026. Three-run warm medians using the Glade Parakeet benchmark:

Audio	Processing	RTFx
20 seconds	0.352 s	56.8×
18 min 15 s	3.937 s	278.2×

Timing includes feature extraction, encoder, decoding, tokenization and native word timing; excludes file reading, chunk planning, initialization, model preparation and warmup. Instruments separately observed ANE predictions in all 21 calls to each encoder/subsampling stage and in batched decoding, with no GPU intervals. Captures use encoder concurrency one for unambiguous stage attribution; throughput runs use two with the same graph functions.

On the full JFK benchmark recording, Whisper-normalized WER is 2.34% (52/2220 errors). All source token IDs match on all 21 identical chunk cuts, each with its own leading silence; the 20-second check also matches stock Transformers generation. This is a limited conversion check on one English recording, not a broad accuracy ranking or independent verification of the publisher's leaderboard claims.

First observed prepare() took 103.7 seconds; subsequent cached preparation in the full-speech report took 0.113 seconds. The cache was not cleared to establish a clean-device compilation estimate. Retained assets were read from network storage; preparation is separate from warm inference. Peak sampled client-plus-separately-attributed-ANE footprint during the full warm runs was 1.32 GB, using nonoverlapping task VM counters. These attribution counters do not establish physical residency or capture every shared driver allocation.

iPhone qualification

iPhone Duo (A20 Pro), iOS 27.1 (24A94425), Release Glade Benchmark app. The full speech took 2.347 seconds (466.8× real time), the median of three uninterrupted foreground runs at nominal thermal state. Every token and chunk boundary matches the source control. Phone timing includes bounded WAV reading and chunk planning in addition to the inference stages above. Peak sampled client footprint was 1.24 GB, excluding uncharged accelerator/driver allocations. First observed preparation took 195.0 seconds, excluded from inference. A separate full-speech Instruments capture observes ANE predictions in every encoder/subsampling call and in batched decoding, with no target GPU intervals. Attribution capture uses encoder concurrency one; throughput uses two.