Update model card: add transcribe_cpp metadata block (raw WER/RTF + capabilities)

README.md

@@ -17,6 +17,24 @@ tags:
   - audio-llm
   - fastconformer
   - qwen3
+transcribe_cpp:
+  wer_librispeech_test_clean:
+    bf16: 1.63
+    f16: 1.63
+    q8_0: 1.63
+    q6_k: 1.63
+    q5_k_m: 1.63
+    q4_k_m: 1.63
+  rtf_m4_max:
+    metal: 39.8
+    cpu: 7.3
+  rtf_ryzen_4750u:
+    vulkan: 4.1
+    cpu: 2.1
+  streaming: false
+  translate: false
+  lang_detect: false
+  timestamps: none
 ---
 
 # canary-qwen-2.5b: transcribe.cpp GGUF
@@ -60,9 +78,8 @@ GGUF presets land at exactly 1.63% (`+0.02` over our reference run,
 same CI band). Investigation of the worst per-utterance differences
 shows scattered token-level noise consistent with BF16 weight
 precision (homophones, word-boundary flips, function-word
-substitutions); see the transcribe.cpp model page for the full
-teacher-forced logit comparison. Reproduce with `scripts/wer/run.py`
-+ `scripts/wer/score.py`.
+substitutions). Reproduce with `scripts/wer/run.py` +
+`scripts/wer/score.py`.
 
 
 ## Usage