From 5b89ef2c6e5895b168f8f150ddce345dcee6be91 Mon Sep 17 00:00:00 2001 From: Alexander Korolev Date: Tue, 1 Jun 2021 11:06:35 +0200 Subject: [PATCH 1/2] fix speaker-embeddings dimension during inference --- TTS/tts/models/tacotron2.py | 1 + 1 file changed, 1 insertion(+) diff --git a/TTS/tts/models/tacotron2.py b/TTS/tts/models/tacotron2.py index fded8f87..a5db64e9 100644 --- a/TTS/tts/models/tacotron2.py +++ b/TTS/tts/models/tacotron2.py @@ -277,6 +277,7 @@ class Tacotron2(TacotronAbstract): if self.num_speakers > 1: if not self.embeddings_per_sample: speaker_embeddings = self.speaker_embedding(speaker_ids)[:, None] + speaker_embeddings = torch.unsqueeze(speaker_embeddings, 0).transpose(1, 2) encoder_outputs = self._concat_speaker_embedding(encoder_outputs, speaker_embeddings) mel_outputs, alignments, stop_tokens = self.decoder.inference_truncated(encoder_outputs) From c1eb9bdccacfb09356282557bae21885ecaa0dfa Mon Sep 17 00:00:00 2001 From: Alexander Korolev Date: Tue, 1 Jun 2021 15:15:26 +0200 Subject: [PATCH 2/2] fix speaker dim inference --- TTS/tts/models/tacotron2.py | 1 + 1 file changed, 1 insertion(+) diff --git a/TTS/tts/models/tacotron2.py b/TTS/tts/models/tacotron2.py index a5db64e9..525eb8b3 100644 --- a/TTS/tts/models/tacotron2.py +++ b/TTS/tts/models/tacotron2.py @@ -255,6 +255,7 @@ class Tacotron2(TacotronAbstract): if self.num_speakers > 1: if not self.embeddings_per_sample: speaker_embeddings = self.speaker_embedding(speaker_ids)[:, None] + speaker_embeddings = torch.unsqueeze(speaker_embeddings, 0).transpose(1, 2) encoder_outputs = self._concat_speaker_embedding(encoder_outputs, speaker_embeddings) decoder_outputs, alignments, stop_tokens = self.decoder.inference(encoder_outputs)