Commit History
Yesterday
-
Sync card from manifest.json
Verified signature23c0cdd
2 days ago
-
Sync card from manifest.json
Verified signature6062ae8
6 days ago
-
Add onnx-fp16: ONNX export for desktop runtimes (split mel, raw-logit decoder)
Verified signature12802bf
9 days ago
-
Sync card from manifest.json
Verified signature8e34d46 -
Card: the blocks the manifest sync maintains
Verified signature7ecb18c -
Weights v0.4.0
Verified signature0991d35
12 days ago
-
Sync card from manifest.json
Verified signatureae825b7
30 days ago
-
Rewrite the card in plain language
Verified signaturee11906f -
Accuracy: plain wording on what error rate to expect
Verified signature7ade41c -
Intro and Try it: plain description, all platforms
Verified signature5310228 -
Speed: per-device table from the Swift SDK 3.5.0 run
Verified signature57fb35f
-
Decoder with narrow decode steps
Verified signaturedecoder.mlmodelc is now a multifunction model: the 16-lane decode step as main, and the same step at 8, 4, 2 and 1 lanes as decoder_8 ... decoder_1. int8 per-channel weights, shared across the functions (12 MB, was 15 MB). Built by voz-training's build_voz.py. Needs a desert-ant-core with narrow-step decoding; an older SDK loads main and runs as before.20e1118
-
Web bundle: the ONNX export the browser SDK loads, 4-bit encoder
Verified signaturea6acd58
-
Rename the LiteRT graphs to their roles
Verified signatureArtifact names describe roles rather than the network behind them, matching the Core ML files, so replacing the recogniser is a new upload rather than an SDK change.fab0da8 -
Move LiteRT bundle to root, matching the cross-platform layout
Verified signaturevocab.json and embedding.f16 are byte-identical to the Core ML ones and stay shared; the LiteRT shape contract goes in meta.litert.json since meta.json carries Core ML-specific values (decode_width 8, two encoder chunks).cfe98f5 -
Add LiteRT (.tflite) bundle for Android under litert/
Verified signature4ec9a70
-
Sync card from manifest.json
Verified signature78e93e1 -
Sync card from manifest.json
Verified signature8914d7b
-
Restore the card after the sync flattened it again
Verified signatureca2eae2
-
Sync card from manifest.json
Verified signaturece46c49 -
Model card: what Voz is, how it is built, and what it scores
Verified signature9228c54
-
Sync card from manifest.json
Verified signature8e7ab8c -
Rename to Voz
Verified signaturef52c67e -
Rename to Voz
Verified signaturedbf7b65
-
Compare against Whisper large-v3-turbo only
Verified signature2808eed
-
Compare against Whisper on the Open ASR Leaderboard
Verified signatureSimplifies the table to the comparison that matters to a reader choosing a model, and corrects it. The previous version compared results from the cleaned dataset configurations against figures from the original ones, which are systematically higher, and so understated the gap. Whisper's numbers here are the leaderboard's own, on the same configurations. 7.40% averaged over six sets, against 6.89% for Whisper large-v3 and 7.00% for large-v3-turbo. Better on meetings by roughly two points, behind on prepared and read speech, at 467 MB entirely on the Neural Engine.9af2c22 -
Scribe: on-device speech recognition for Apple platforms
d42390b