fix vocoder train

cosyvoice/flow/flow.py

@@ -91,7 +91,7 @@ class MaskedDiffWithXvec(torch.nn.Module):
         conds = conds.transpose(1, 2)
 
         mask = (~make_pad_mask(feat_len)).to(h)
-        # NOTE 这一句应该是不需要的，应该h已经过length_regulator跟feat一样的shape了
+        # NOTE this is unnecessary, feat/h already same shape
         feat = F.interpolate(feat.unsqueeze(dim=1), size=h.shape[1:], mode="nearest").squeeze(dim=1)
         loss, _ = self.decoder.compute_loss(
             feat.transpose(1, 2).contiguous(),
@@ -117,7 +117,7 @@ class MaskedDiffWithXvec(torch.nn.Module):
         embedding = F.normalize(embedding, dim=1)
         embedding = self.spk_embed_affine_layer(embedding)
 
-        # concat text and prompt_text
+        # concat speech token and prompt speech token
         token_len1, token_len2 = prompt_token.shape[1], token.shape[1]
         token, token_len = torch.concat([prompt_token, token], dim=1), prompt_token_len + token_len
         mask = (~make_pad_mask(token_len)).unsqueeze(-1).to(embedding)

cosyvoice/flow/length_regulator.py

@@ -51,6 +51,7 @@ class InterpolateRegulator(nn.Module):
 
     def inference(self, x1, x2, mel_len1, mel_len2, input_frame_rate=50):
         # in inference mode, interploate prompt token and token(head/mid/tail) seprately, so we can get a clear separation point of mel
+        # NOTE 20 corresponds to token_overlap_len in cosyvoice/cli/model.py
         # x in (B, T, D)
         if x2.shape[1] > 40:
             x2_head = F.interpolate(x2[:, :20].transpose(1, 2).contiguous(), size=int(20 / input_frame_rate * 22050 / 256), mode='linear')