Update README.md

README.md

@@ -183,11 +183,22 @@ The Low Frame-rate Speech Codec is trained on a total of 28.7k hrs of speech dat
 
 We evaluated our codec using multiple objective audio quality metrics across two distinct test sets. Additionally, we compared our model's performance with state-of-the-art codecs. For further details, please refer to [our paper](https://arxiv.org/abs/2409.12117).
 
+Please note that the released checkpoint yields slightly different results compared to those reported in the paper. Due to legal data constraints, we retrained the model after removing one speaker from the training set. This retraining was performed for 170k steps, compared to the original 124k steps, leading to slight improvements across almost all metrics.
+
+Paper results:
+
 | Dataset     | Squim MOS (↑)     |SI-SDR(↑)      |Mel Dist. (↓)      |STFT Dist.(↓) | CER (↓)|
 |:-----------:|:----------:|:----------:|:----------:|:-----------:|:-----------:|
 | MLS | 4.43       | 4.46      | 0.147      | 0.061       | 2.09       | 
 | DAPS | 4.68       | 6.93       | 0.142       | 0.058       | 0.86      |
 
+Released checkpoint results:
+
+| Dataset     | Squim MOS (↑)     |SI-SDR(↑)      |Mel Dist. (↓)      |STFT Dist.(↓) | CER (↓)|
+|:-----------:|:----------:|:----------:|:----------:|:-----------:|:-----------:|
+| MLS |  4.43     |  4.77   |    0.143  |  0.060     |     2.16  | 
+| DAPS | 4.69       | 8.07       | 0.136       | 0.056      |    0.77  |
+
 
 
 ## Software Integration