virgil-coach/detailkeeper-runtime

Detail Keeper runtime

v0.7.0. The models Detail Keeper, by Virgil, runs on a Mac, ready to load: most of what the app downloads on its first start, so that it needs no account and builds nothing. They are in the layouts and precisions the app's engine loads (MLX on Apple silicon); they are not drop-in files for transformers. Apache 2.0 (LICENSE); NOTICE lists the source, authors, citation and changes of each model.

Folder	What
writer/	qwen3-asr-1.7b-own-names-v0.7.0-audio-text-note-affine8.safetensors: Qwen3-ASR-1.7B with the own-names adapter v0.7.0 (audio-text-note) merged in; and tokenizer.json, Qwen3-ASR-1.7B's tokenizer
wav2vec2-lv-60-espeak-cv-ft/	the spotter's phoneme recognizer, facebook/wav2vec2-lv-60-espeak-cv-ft, float16

The spotter's other model, microsoft/wavlm-base-plus, is not here: the app fetches it from Microsoft's repo (commit 98fd61b9, full precision), under its own licence, CC BY-SA 3.0 (microsoft/UniSpeech). With it, a first start downloads 3.62 GB.

The writer

The adapter (v0.7.0: LoRA, rank 64, alpha 128, on the text decoder's attention and MLP projections) is merged into Qwen3-ASR-1.7B's bf16 weights (mlx-community/Qwen3-ASR-1.7B-bf16, commit e1f6c266, the same numbers as Qwen's checkpoint in MLX's layout) in float32. Then the text decoder's projections and its embedding (which is also the output layer) are quantized to 8 bits, MLX affine with group size 64; the audio encoder stays bf16. Each adapted projection also keeps the adapter's two small matrices (lora.<weight>.a, and .b scaled by alpha / r), so the engine can take the adapter's share back out for one call and run the plain recognizer from the same weights. The file is what the engine builds from the adapter on a Mac, bit for bit.

How the writer takes its shortlist, and what it gets right and wrong, is on the adapter's card (v0.7.0). The adapter was trained on public-domain, CC0, CC BY 3.0 and CC BY 4.0 speech and noise (LibriVox, LibriTTS, YODAS, VoxPopuli, LibriTTS-R, NOTSOFAR-1, MUSAN, FSD50K), and on synthetic speech in 69 adult LibriTTS readers' voices cloned with Qwen3-TTS 1.7B (Apache 2.0); its attribution/ folder lists every source, its licence and every item used, with the credit each CC BY item asks for.

English only. The engine forces the writer's language to English, and the adapter was trained and tested on English speech alone (see its card). It does not write other languages.

The phoneme recognizer

It runs in float32. Its weights are stored here in float16; the engine converts them back to float32 as it loads them, so the only change is the weights' rounding. masked_spec_embed (used only in pre-training) is left out. The rest is unchanged, config files included.

Checked against full precision

The spotter run with this float16 phoneme recognizer and with the original float32 one (WavLM at full precision in both), on 1,104 windows of eight public-domain audiobook chapters (LibriVox):

Lexicon	Same top ten (and in the same order)	Recall@10 of names said, float32 / float16
each chapter's own taught names	1,104 of 1,104 windows	122 / 122 of 122
758 names	1,104 of 1,104 windows	247 / 247 of 249

No window's shortlist changed.

SHA-256

d7109615f66dd8ee9df9222ae1bbf61f9f9b2fdcae5996ef680bae64842f6162  writer/qwen3-asr-1.7b-own-names-v0.7.0-audio-text-note-affine8.safetensors
fe1fad59be22a41ee293363fcf95fdedbc7c93f3b49270b1d2e18bd1399a7a05  writer/tokenizer.json
682db57c7150fcaacb78599e4e92751cae0d4d17e36bfecd1a5d0b7aa7654f79  wav2vec2-lv-60-espeak-cv-ft/model.safetensors