Sin2pi commited on
Commit
eaeb462
·
verified ·
1 Parent(s): 3c6d395

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +10 -0
README.md CHANGED
@@ -23,6 +23,16 @@ tags:
23
 
24
  ASR model + pitch aware relative positional embeddings.
25
 
 
 
 
 
 
 
 
 
 
 
26
  <img width="1363" height="732" alt="pitch_spectrogram" src="https://github.com/user-attachments/assets/ceb65e94-7df4-41b7-aa3d-c4aa4c6c0717" />
27
 
28
  <img width="233" height="77" alt="legend" src="https://github.com/user-attachments/assets/fad84550-a199-43b3-8471-d011a9fd6f94" />
 
23
 
24
  ASR model + pitch aware relative positional embeddings.
25
 
26
+ ### Decrease WER significantly compared to standard inverse frequency. 'eval_wer': 35.3
27
+
28
+ def _compute_freqs_base(self):
29
+ mel_scale = torch.pow(10, torch.linspace(0, 2595 * torch.log10(torch.tensor(1 + 4000/200)), self.head_dim // 2, device=device, dtype=dtype) / 2595) - 1
30
+ return 200 * mel_scale / 1000
31
+
32
+ ### Standared inv freqs: 'eval_wer': 61.6
33
+ freqs = 1.0 / (self.theta ** (torch.arange(0, self.head_dim, 2, device=device, dtype=dtype) / (self.head_dim // 2)))
34
+
35
+
36
  <img width="1363" height="732" alt="pitch_spectrogram" src="https://github.com/user-attachments/assets/ceb65e94-7df4-41b7-aa3d-c4aa4c6c0717" />
37
 
38
  <img width="233" height="77" alt="legend" src="https://github.com/user-attachments/assets/fad84550-a199-43b3-8471-d011a9fd6f94" />