Skip to content

SOTA Open Source TTS

Fix inference without attention mask

Lengyue254e93e63287
diff --git a/speech_lm/models/whisper_vq.py b/speech_lm/models/whisper_vq.py
index 28c70f6..f1721ae 100644
--- a/speech_lm/models/whisper_vq.py
+++ b/speech_lm/models/whisper_vq.py
@@ -90,8 +90,8 @@ class WhisperVQ(nn.Module):
if attention_mask is not None:
assert attention_mask.ndim == 2, "Attention mask must be 2D"
- # Whisper will downsample by 2
- attention_mask = attention_mask[:, ::2]
+ # Whisper will downsample by 2
+ attention_mask = attention_mask[:, ::2]
with torch.no_grad():
hidden_states = self.whisper.model.encoder(