Newer interface

.gitattributes

@@ -25,3 +25,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zstandard filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text

hyperparams.yaml

@@ -1,6 +1,12 @@
+n_mels: 40
+
 feature_extractor: !new:speechbrain.lobes.features.Fbank
     n_fft: 400
-    n_mels: 40
+    n_mels: !ref <n_mels>
+
+feature_scaler: !new:custom.FeatureScaler
+    num_in: !ref <n_mels>
+    scale: 0.5
 
 normalizer: !new:speechbrain.processing.features.InputNormalization
     norm_type: global

interface.py

@@ -1,6 +1,14 @@
 import torch
 import speechbrain as sb
 
+class FeatureScaler(torch.nn.Module):
+    def __init__(self, num_in, scale):
+        super().__init__()
+        self.scaler = torch.nn.eye(num_in) * scale
+
+    def forward(x):
+        return x * self.scaler
+
 class Custom(sb.pretrained.interfaces.Pretrained):
     MODULES_NEEDED = ["normalizer"]
     HPARAMS_NEEDED = ["feature_extractor"]
@@ -8,7 +16,8 @@ class Custom(sb.pretrained.interfaces.Pretrained):
     def feats_from_audio(self, audio, lengths=torch.tensor([1.0])):
         feats = self.hparams.feature_extractor(audio)
         normalized = self.mods.normalizer(feats, lengths)
-        return normalized
+        scaled = self.mods.feature_scaler(normalized)
+        return scaled
 
     def feats_from_file(self, path):
         audio = self.load_audio(path)