Benchmarks

All done with 200s chunks, M1 Macbook Air

Inference speed

model CPU fp32 CPU fp16 GPU fp32 GPU fp16 fast_v1 3679 +-165 - 7084 +-311 - balanced_v1 1940 +- 65 - 5354 +- 9 - heavy_v1 403 +- 9 402 +- 7 1292 +- 5 1870 +- 21 model_general_v3 997 +- 2 981 +- 3 2866 +- 11 5377 +- 26

Analysis speed

288 files, ~24h total, .flac from ssd (even sample data)

fast: 4,800, 4,800 (fast decoding) balanced: 3,300 heavy: 942x (full precision), 1,454 (half precision)

4 files, ~5.5 days

fast: 5,000x balanced: 3,500 heavy (fp16): 2,020x (faster than benchmark??) model_general_v3: 4,100 (early steady state was ~4,500?)