--conv-frontend=./sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25/conv_frontend.onnx \\ --encoder=./sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25/encoder.int8.onnx ...
The encoder employs a DenseNet-B (bottleneck) architecture with three dense blocks separated by transition layers. Each bottleneck layer consists of a 1x1 convolution (expanding to 4x growth rate) ...