tested gl/wavglow working

taco2/stft.py

@@ -84,8 +84,8 @@ class STFT(torch.nn.Module):
             forward_basis *= fft_window
             inverse_basis *= fft_window
 
-        self.register_buffer("forward_basis", forward_basis.float())
-        self.register_buffer("inverse_basis", inverse_basis.float())
+        self.register_buffer("forward_basis", forward_basis.float().to(DEVICE))
+        self.register_buffer("inverse_basis", inverse_basis.float().to(DEVICE))
 
     def transform(self, input_data):
         num_batches = input_data.size(0)
@@ -121,10 +121,10 @@ class STFT(torch.nn.Module):
         return magnitude, phase
 
     def inverse(self, magnitude, phase):
+        phase = phase.to(DEVICE)
         recombine_magnitude_phase = torch.cat(
             [magnitude * torch.cos(phase), magnitude * torch.sin(phase)], dim=1
         )
-
         inverse_transform = F.conv_transpose1d(
             recombine_magnitude_phase,
             Variable(self.inverse_basis, requires_grad=False),
@@ -144,11 +144,10 @@ class STFT(torch.nn.Module):
             # remove modulation effects
             approx_nonzero_indices = torch.from_numpy(
                 np.where(window_sum > tiny(window_sum))[0]
-            )
+            ).to(DEVICE)
             window_sum = torch.autograd.Variable(
                 torch.from_numpy(window_sum), requires_grad=False
-            )
-            window_sum = window_sum.to(DEVICE)
+            ).to(DEVICE)
             inverse_transform[:, :, approx_nonzero_indices] /= window_sum[
                 approx_nonzero_indices
             ]

taco2/tts.py

@@ -108,7 +108,7 @@ class TTSModel(object):
             with torch.no_grad():
                 audio_t = self.waveglow.infer(mel_outputs_postnet, sigma=0.666)
                 audio_t = self.denoiser(audio_t, 0.1)[0]
-            audio = audio_t[0].data.cpu().numpy()
+            audio = audio_t[0].data
         elif vocoder == "gl":
             mel_decompress = self.taco_stft.spectral_de_normalize(mel_outputs_postnet)
             mel_decompress = mel_decompress.transpose(1, 2).data.cpu()
@@ -116,25 +116,32 @@ class TTSModel(object):
             spec_from_mel = torch.mm(mel_decompress[0], self.taco_stft.mel_basis)
             spec_from_mel = spec_from_mel.transpose(0, 1).unsqueeze(0)
             spec_from_mel = spec_from_mel * spec_from_mel_scaling
+            spec_from_mel = (
+                spec_from_mel.cuda() if torch.cuda.is_available() else spec_from_mel
+            )
             audio = griffin_lim(
                 torch.autograd.Variable(spec_from_mel[:, :, :-1]),
                 self.taco_stft.stft_fn,
                 60,
             )
             audio = audio.squeeze()
-            audio = audio.cpu().numpy()
         else:
             raise ValueError("vocoder arg should be one of [wavglow|gl]")
+        audio = audio.cpu().numpy()
         return audio
 
-    def synth_speech(self, text, speed: 1.0, sample_rate=OUTPUT_SAMPLE_RATE):
+    def synth_speech(
+        self, text, speed: float = 1.0, sample_rate: int = OUTPUT_SAMPLE_RATE
+    ):
         audio = self.synth_speech_array(text, VOCODER_MODEL)
 
         return postprocess_audio(
-            audio, src_rate=self.hparams.sample_rate, dst_rate=sample_rate, tempo=speed
+            audio, src_rate=self.hparams.sampling_rate, dst_rate=sample_rate, tempo=speed
         )
 
-    def synth_speech_fast(self, text, speed: 1.0, sample_rate=OUTPUT_SAMPLE_RATE):
+    def synth_speech_fast(
+        self, text, speed: float = 1.0, sample_rate: int = OUTPUT_SAMPLE_RATE
+    ):
         mel_outputs_postnet = self.generate_mel_postnet(text)
 
         mel_decompress = self.taco_stft.spectral_de_normalize(mel_outputs_postnet)
@@ -152,7 +159,7 @@ class TTSModel(object):
         audio = audio.cpu().numpy()
 
         return postprocess_audio(
-            audio, tempo=speed, src_rate=self.hparams.sample_rate, dst_rate=sample_rate,
+            audio, tempo=speed, src_rate=self.hparams.sampling_rate, dst_rate=sample_rate
         )