Skip to content

SOTA Open Source TTS

Fix quantizer norm

Lengyued28ba743d301
diff --git a/speech_lm/models/hubert_vq.py b/speech_lm/models/hubert_vq.py
index d6db4b9..5e6a9a4 100644
--- a/speech_lm/models/hubert_vq.py
+++ b/speech_lm/models/hubert_vq.py
@@ -46,6 +46,7 @@ class HubertVQ(nn.Module):
param.requires_grad = True
# Quantization
+ self.quantizer_ln = nn.LayerNorm(self.hubert.config.hidden_size)
self.quantizer = VectorQuantization(
codebook_size=codebook_size,
dim=self.hubert.config.hidden_size,
@@ -164,6 +165,7 @@ class HubertVQ(nn.Module):
)
# Quantize
+ hidden_states = self.quantizer_ln(hidden_states)
quantize, _, vq_loss = self.quantizer(hidden_states.transpose(1, 2))
quantize = quantize.transpose(1, 2)