# FastWave.AXERA — 精度报告 ## 量化配置 - Pulsar2 8.0 / AX650 / NPU3,`compiler.check=2` - 数据类型:**U16 激活 / S8 权重 / U16 输出**(`DEFAULT` 全图),`highest_mix_precision=false` - 标定:真实 48 kHz 语音,8/12/16/24 kHz 退化输入,8 个 EDM 采样步全采样(64 个样本) ### 量化工件内精度(Pulsar2 EndToEnd 逐算子) - 统计算子 2245 个,cosine min 0.90369 / mean 0.99962 - 图输出 `y` cosine **0.99982**(MSE 0.00000) ## 单步 NFE 精度(AXMODEL 板端 vs 浮点 ONNX,held-out 数据) | case | 输入采样率 | 采样步 | cosine | SNR (dB) | max abs err | |---|---|---|---|---|---| | 0 | 8000 Hz | 0 | 0.997546 | 23.09 | 0.0108 | | 0 | 8000 Hz | 1 | 0.994159 | 18.68 | 0.0152 | | 0 | 8000 Hz | 2 | 0.995742 | 20.59 | 0.0098 | | 0 | 8000 Hz | 3 | 0.997580 | 22.58 | 0.0072 | | 0 | 8000 Hz | 4 | 0.992172 | 18.07 | 0.0135 | | 0 | 8000 Hz | 5 | 0.999040 | 27.15 | 0.0039 | | 0 | 8000 Hz | 6 | 0.999682 | 31.93 | 0.0045 | | 0 | 8000 Hz | 7 | 0.988324 | 16.25 | 0.0057 | | 1 | 24000 Hz | 0 | 0.999696 | 32.12 | 0.0717 | | 1 | 24000 Hz | 1 | 0.999776 | 33.42 | 0.0647 | | 1 | 24000 Hz | 2 | 0.999697 | 32.18 | 0.0809 | | 1 | 24000 Hz | 3 | 0.999676 | 31.88 | 0.0940 | | 1 | 24000 Hz | 4 | 0.999702 | 32.25 | 0.1098 | | 1 | 24000 Hz | 5 | 0.999865 | 35.69 | 0.1106 | | 1 | 24000 Hz | 6 | 0.999938 | 39.06 | 0.0829 | | 1 | 24000 Hz | 7 | 0.999516 | 30.14 | 0.0062 | | 2 | 16000 Hz | 0 | 0.996778 | 21.91 | 0.0157 | | 2 | 16000 Hz | 1 | 0.990080 | 16.38 | 0.0148 | | 2 | 16000 Hz | 2 | 0.994315 | 19.38 | 0.0092 | | 2 | 16000 Hz | 3 | 0.996483 | 21.25 | 0.0067 | | 2 | 16000 Hz | 4 | 0.986250 | 15.62 | 0.0125 | | 2 | 16000 Hz | 5 | 0.998712 | 25.86 | 0.0035 | | 2 | 16000 Hz | 6 | 0.999681 | 31.91 | 0.0025 | | 2 | 16000 Hz | 7 | 0.983171 | 14.57 | 0.0056 | | 3 | 12000 Hz | 0 | 0.999837 | 34.64 | 0.0651 | | 3 | 12000 Hz | 1 | 0.999933 | 38.65 | 0.0465 | | 3 | 12000 Hz | 2 | 0.999946 | 39.01 | 0.0457 | | 3 | 12000 Hz | 3 | 0.999945 | 39.10 | 0.0370 | | 3 | 12000 Hz | 4 | 0.999943 | 39.40 | 0.1066 | | 3 | 12000 Hz | 5 | 0.999955 | 40.30 | 0.1270 | | 3 | 12000 Hz | 6 | 0.999983 | 44.69 | 0.0420 | | 3 | 12000 Hz | 7 | 0.999934 | 38.81 | 0.0060 | - cosine 最小值 **0.983171**,平均值 0.997096 - 判定:PASS(端到端验证见下,quiet chunk 单步略低) ## 端到端精度(一段 1.4 s 8 kHz 输入) - NPU 输出 vs 浮点输出:cosine **0.99842**,SNR 24.93 dB ## NPU 延迟 - 单次 NFE:平均 499.21 ms(min 498.57 / max 499.94) - 8 步 / chunk(0.68 s):约 3.99 s