Addressing Ning's comments.

src/seamless_communication/assets/cards/xlsr2_1b_v2.yaml

@@ -7,4 +7,4 @@
 name: xlsr2_1b_v2
 model_type: wav2vec2
 model_arch: xlsr2_1b_v2
-checkpoint: "file://private/home/changhan/data/models/wav2vec2/xlsr2_1b_v2.pt"
+checkpoint: "https://dl.fbaipublicfiles.com/seamlessM4T/models/unit_extraction/xlsr2_1b_v2.pt"

src/seamless_communication/models/unit_extraction/wav2vec2_layer_output.py

@@ -56,8 +56,8 @@ def _encoder_xlsr2_1b_v2() -> Wav2Vec2EncoderConfig:
         num_encoder_layers=48,
         num_encoder_attn_heads=16,
         ffn_inner_dim=5120,
-        dropout_p=0.0,
-        attn_dropout_p=0.0,
+        dropout_p=0.1,
+        attn_dropout_p=0.1,
         layer_drop_p=0.0,
         norm_order=TransformerNormOrder.PRE,
         depthwise_conv_kernel_size=0,