Update README.md

README.md

@@ -21,7 +21,9 @@ tags:
 
 ---
 
-ASR model that uses audio frequencies instead of spectrograms. + pitch aware relative positional embeddings. 
+ASR model that uses audio frequencies instead of spectrograms + pitch aware relative positional embeddings. 
+
+
 
 Questions:
 
@@ -35,8 +37,6 @@ Questions:
 
 
 
-<img width="780" alt="cc5" src="https:github.comuser-attachmentsassets106ebe75-f1db-4f85-bdae-818b114fedd2"  >
-
 
 To explore the relationship between pitch and rotary embeddings, the model implements three complementary pitch based enhancements:
 
@@ -44,6 +44,16 @@ To explore the relationship between pitch and rotary embeddings, the model imple
 2. Direct similarity bias: A pitch based similarity bias is added directly to the attention mechanism.
 3. Variable radii in torch.polar: The unit circle radius 1.0 in the torch.polar calculation is replaced with variable radii derived from f0. This creates acoustically-weighted positional encodings, so each position in the embedding space reflects the acoustic prominence in the original speech. This approach effectively adds phase and amplitutde information without significant computational overhead.
 
+4. Initial findings suggest that f0 is a superior input to spectrograms for ASR models cutting training time almost by half without negatively effecting other metrics such as WER or loss.
+   
+
+<img width="1816" height="707" alt="pitchee2" src="https://github.com/user-attachments/assets/17c89ebf-3373-4dd5-b510-95fa96774ec1" />
+
+
+
+
+
+
 The function `torch.polar` constructs a complex tensor from polar coordinates:
 
 ````python
@@ -296,3 +306,5 @@ MaxFactor is a custom PyTorch optimizer with adaptive learning rates and special
 
 ** this model deviates in a lot of ways from standard transformer models.
 
+
+