subthx/kitsune-onnx

Kitsune transcribe (ja) for subthx: ONNX

Kitsune ja transcription models (0.05B, 0.1B, 0.3B) as ONNX for subthx. The weights are from Multy123's Kitsune (ParakeetForCTC, CC BY-NC 4.0); the ONNX graphs with argmax inside are exported by subthx (tools/export_kitsune.py). Copy the repo's contents into the models folder of subthx.

Files

file	precision / quantization
kitsune-p0.05b/config.json	-
kitsune-p0.05b/tokenizer.json	-
kitsune-p0.05b/tokenizer_config.json	-
kitsune-p0.05b/model.int8.onnx	int8 matmuls, fp32 convolutions (CPU)
kitsune-p0.05b/model.mixed16b.onnx	mixed precision: fp16 weights, fp32 attention scores (GPU, CUDA)
kitsune-p0.1b/config.json	-
kitsune-p0.1b/tokenizer.json	-
kitsune-p0.1b/tokenizer_config.json	-
kitsune-p0.1b/model.int8.onnx	int8 matmuls, fp32 convolutions (CPU)
kitsune-p0.1b/model.mixed16b.onnx	mixed precision: fp16 weights, fp32 attention scores (GPU, CUDA)
kitsune-p0.3b/config.json	-
kitsune-p0.3b/tokenizer.json	-
kitsune-p0.3b/tokenizer_config.json	-
kitsune-p0.3b/model.int8.onnx	int8 matmuls, fp32 convolutions (CPU)
kitsune-p0.3b/model.mixed16b.onnx	mixed precision: fp16 weights, fp32 attention scores (GPU, CUDA)
nemo80_fbanks.npy	-

Sizes and sha256: manifest.json.