| { | |
| "name": "wakeup_nihao_aixin_ax620e", | |
| "framework": "ONNX", | |
| "task": "wake_word_detection", | |
| "route": "general", | |
| "inputs": [ | |
| { | |
| "name": "fbank", | |
| "shape": [ | |
| 1, | |
| 26, | |
| 64 | |
| ], | |
| "dtype": "float32", | |
| "layout": "NCHW" | |
| } | |
| ], | |
| "outputs": [ | |
| { | |
| "name": "logits", | |
| "shape": [ | |
| 1, | |
| 2, | |
| 64 | |
| ], | |
| "dtype": "float32" | |
| } | |
| ], | |
| "preprocess": "reference log-fbank: sample/32768 -> hann(512,periodic) -> rfft magnitude -> mel filterbank (26) -> log(x+1e-7) -> Q6.10 (q=x*1024/1024)", | |
| "postprocess": "wake word '你好,爱芯': wake score = logits[1, last]; 3-frame sliding sum vs threshold (0.608 DET_MODE_90 / 0.615 DET_MODE_95)", | |
| "sdk_interface": { | |
| "input": "int16 pcm mono 16k, 512 samples per frame", | |
| "output": "per-frame wake score + trigger event" | |
| }, | |
| "hardware": "AX620E", | |
| "npu_mode": "NPU2", | |
| "training": "fine-tuned from WakeNet9 with edge-tts data (see model_convert/train/)" | |
| } | |