Model card: list the XNNPACK int8 variants and their accuracy

README.md

@@ -22,6 +22,18 @@ Upstream model: [whisper-base](https://huggingface.co/openai/whisper-base)
 | `vulkan/whisper_base_vulkan_fp16.pte` | vulkan | fp16 |
 | `vulkan/whisper_base_vulkan_int8.pte` | vulkan | int8 |
 | `xnnpack/whisper_base_xnnpack_fp32.pte` | xnnpack | fp32 |
+| `xnnpack/whisper_base_xnnpack_int8.pte` | xnnpack | int8 |
+
+## XNNPACK int8
+
+`xnnpack/whisper_base_xnnpack_int8.pte` is 8da8w (int8 dynamic activations, int8
+per-channel weights). Greedy WER (CER for zh) against the fp32 file, LibriSpeech
+test-clean (150 clips) and FLEURS (80 per language):
+
+| | en | de | es | pl | fr | zh |
+| --- | --- | --- | --- | --- | --- | --- |
+| fp32 | 4.12 | 24.88 | 13.66 | 47.35 | 40.76 | 21.27 |
+| int8 | 3.94 | 25.04 | 13.87 | 46.39 | 41.09 | 21.27 |
 
 ## Repository structure
 
@@ -38,6 +50,7 @@ vulkan/whisper_base_vulkan_fp16.pte    335 MB
 vulkan/whisper_base_vulkan_int8.pte    196 MB
 xnnpack/config.json                    1.4 kB
 xnnpack/whisper_base_xnnpack_fp32.pte  380 MB
+xnnpack/whisper_base_xnnpack_int8.pte  254 MB
 ```
 
 ## Compatibility