Update README.md

README.md

@@ -21,8 +21,11 @@ tags:
 
 ---
 
+
 ASR model + pitch aware relative positional embeddings. 
 
+<img width="1431" height="636" alt="pitch_spectrogram" src="https://github.com/user-attachments/assets/2db80884-7e27-4a24-ad38-c9b8c28f26da" />
+<img width="233" height="77" alt="legend" src="https://github.com/user-attachments/assets/fad84550-a199-43b3-8471-d011a9fd6f94" />
 
 
 Questions:
@@ -36,8 +39,6 @@ Questions:
 ---
 
 
-
-
 To explore the relationship between pitch and rotary embeddings, the model implements three complementary pitch based enhancements:
 
 1. Pitch modulated theta Pitch f0 is used to modify the theta parameter, dynamically adjusting the rotary frequency.
@@ -47,7 +48,6 @@ To explore the relationship between pitch and rotary embeddings, the model imple
 
 
 
-
 The function `torch.polar` constructs a complex tensor from polar coordinates:
 
 ````python
@@ -66,7 +66,7 @@ Here are the abbreviated steps for replacing theta and radius in the rotary forw
 
 ```python
 
-f0 = f0.todevice, dtype # feature extracted during processing
+
 if f0 is not None:
     if f0.dim == 2:
         f0 = f0.squeeze0 
@@ -220,19 +220,14 @@ The Complex Frequency Result:
       Transitions: Natural pitch changes
 
 
-<img width="780" alt="cc4" src="https:github.comuser-attachmentsassets165a3f18-659a-4e2e-a154-a3456b667bae"  >
 
 
 
-----
 
-This model sometimes uses :
+----
 
-https:github.comsine2piMaxfactor
 
-MaxFactor is a custom PyTorch optimizer with adaptive learning rates and specialized handling for matrix parameters.
 
-** model deviates from standard transformer models.