SOTA Open Source TTS
Fix speaker encoder
Lengyue77371cc16b6f
diff --git a/fish_speech/models/vqgan/lit_module.py b/fish_speech/models/vqgan/lit_module.pyindex 2d364b8..cbf9243 100644--- a/fish_speech/models/vqgan/lit_module.py+++ b/fish_speech/models/vqgan/lit_module.py@@ -105,7 +105,7 @@ class VQGAN(L.LightningModule):(z_q_audio, z_p),(m_p_text, logs_p_text),(m_q, logs_q),- ) = self.generator(features, feature_lengths, gt_mels, feature_lengths)+ ) = self.generator(features, feature_lengths, gt_mels)y_hat_mel = self.mel_transform(y_hat.squeeze(1))y_mel = slice_segments(gt_mels, ids_slice, self.segment_size // self.hop_length)diff --git a/fish_speech/models/vqgan/modules/encoders.py b/fish_speech/models/vqgan/modules/encoders.pyindex a61edf6..50c597c 100644--- a/fish_speech/models/vqgan/modules/encoders.py+++ b/fish_speech/models/vqgan/modules/encoders.py@@ -202,6 +202,6 @@ class SpeakerEncoder(nn.Module):for block in self.blocks:x = block(x, x, x, key_padding_mask=x_mask)[0]- x = self.out_proj(x[:, 0])+ x = self.out_proj(x[:, :1, :]).mTreturn x