Commit History
2 days ago
-
README: the encoder at the GPU's default precision, and what the NPU's 10 s windows cost
Verified signaturePerformance: the 30 s encoder file at the GPU's default precision against fp32 (Galaxy S26 156.7 against 344.6 ms per window, alternating in one session; Pixel 8a 581.6 against 1,084.9 ms) and the 54-minute recording's word error rate with the reader at fp32 (3.792 % on both phones, against 3.812 %). The NPU option's extra errors come from its 10 s windows (427 errors at float32 in the same windows, 428 on the NPU). Known weaknesses: float16 in the encoder keeps the word error rate, in the reader it changes words. The model files are unchanged.b0fccaa
8 days ago
-
Card: the browser demo Space, linked next to the Android app
Verified signature1888114
13 days ago
-
Card: the Galaxy S26 NPU option with the reader (encoder on the NPU in 10 s windows, reader on GPU fp32 on its own thread): the 54-minute recording in 71.1 s, 45.8x real time, WER 4.328 %
Verified signature84f653a
18 days ago
-
Reader v2 (6 layers, 19.6 M, LibriSpeech 960 h): the two reader graphs and FILES.sha256; card with the v2 numbers on the Pixel 8a, the Galaxy S26 rows, and the reader repository's v2 revision
Verified signatureaf9fe50 -
Card: Conversion section (sources and revisions, tool versions, the two commands, the byte-identical rebuild of 2026-10-15)
Verified signature80106e1
19 days ago
-
EmbeddingGemma 2 Transcribe for LiteRT: fp16-safe encoder (fp16 weights) for 30 s and 10 s windows, reader v1 graphs, vocab, card
Verified signature54d0e90 -
initial commit
Verified signature175b197