relax scipy version

2. updated path variable

setup.py

@@ -12,15 +12,22 @@ with open("HISTORY.rst") as history_file:
 
 requirements = [
     "klepto==0.1.6",
-    "numpy==1.16.4",
+    "numpy~=1.16.4",
     "inflect==0.2.5",
     "librosa==0.6.0",
-    "scipy==1.3.0",
+    "scipy~=1.3",
     "Unidecode==1.0.22",
-    "torch==1.1.0",
-    "PyAudio==0.2.11"
+    "torch~=1.1.0",
 ]
 
+extra_requirements = {
+    "playback": ["PyAudio==0.2.11"],
+    "server": [
+        "google-cloud-texttospeech==1.0.1",
+        "rpyc==4.1.4",
+    ],
+}
+
 setup_requirements = ["pytest-runner"]
 
 test_requirements = ["pytest"]
@@ -44,6 +51,7 @@ setup(
     ],
     description="Taco2 TTS package.",
     install_requires=requirements,
+    extras_require=extra_requirements,
     long_description=readme + "\n\n" + history,
     include_package_data=True,
     keywords="tacotron2 tts",
@@ -53,7 +61,12 @@ setup(
     test_suite="tests",
     tests_require=test_requirements,
     url="https://github.com/malarinv/tacotron2",
-    version="0.2.0",
+    version="0.3.0",
     zip_safe=False,
-    entry_points={"console_scripts": ("tts_debug = taco2.tts:main",)},
+    entry_points={
+        "console_scripts": (
+            "tts_debug = taco2.tts:main",
+            "tts_rpyc_server = taco2.server.__main__:main",
+        )
+    },
 )

taco2/hparams.py

@@ -35,7 +35,7 @@ class HParams(object):
     # Audio Parameters             #
     ################################
     max_wav_value = 32768.0
-    sampling_rate = 16000
+    sampling_rate = 22050
     filter_length = 1024
     hop_length = 256
     win_length = 1024

taco2/server/__main__.py

@@ -0,0 +1,48 @@
+import os
+import logging
+
+import rpyc
+from rpyc.utils.server import ThreadedServer
+
+from .backend import TTSSynthesizer
+
+
+tts_backend = os.environ.get("TTS_BACKEND", "taco2")
+tts_synthesizer = TTSSynthesizer(backend=tts_backend)
+
+
+class TTSService(rpyc.Service):
+    def on_connect(self, conn):
+        # code that runs when a connection is created
+        # (to init the service, if needed)
+        pass
+
+    def on_disconnect(self, conn):
+        # code that runs after the connection has already closed
+        # (to finalize the service, if needed)
+        pass
+
+    def exposed_synth_speech(self, utterance: str):  # this is an exposed method
+        speech_audio = tts_synthesizer.synth_speech(utterance)
+        return speech_audio
+
+    def exposed_synth_speech_cb(
+        self, utterance: str, respond
+    ):  # this is an exposed method
+        speech_audio = tts_synthesizer.synth_speech(utterance)
+        respond(speech_audio)
+
+
+def main():
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
+    )
+    port = int(os.environ.get("TTS_RPYC_PORT", "7754"))
+    logging.info("starting tts server...")
+    t = ThreadedServer(TTSService, port=port)
+    t.start()
+
+
+if __name__ == "__main__":
+    main()

taco2/server/backend.py

@@ -0,0 +1,45 @@
+import os
+
+from google.cloud import texttospeech
+from ..tts import TTSModel
+
+
+tts_model_weights = os.environ.get(
+    "TTS_MODELS", "models/tacotron2_statedict.pt,models/waveglow_256channels.pt"
+)
+
+tts_creds = os.environ.get(
+    "GOOGLE_APPLICATION_CREDENTIALS", "/code/config/gre2e/keys/gre2e_gcp.json"
+)
+taco2, wav_glow = tts_model_weights.split(",", 1)
+
+
+class TTSSynthesizer(object):
+    """docstring for TTSSynthesizer."""
+
+    def __init__(self, backend="taco2"):
+        super(TTSSynthesizer, self).__init__()
+        if backend == "taco2":
+            tts_model = TTSModel(f"{taco2}", f"{wav_glow}")  # Loads the models
+            self.synth_speech = tts_model.synth_speech
+        elif backend == "gcp":
+            client = texttospeech.TextToSpeechClient()
+            # Build the voice request, select the language code ("en-US") and the ssml
+            # voice gender ("neutral")
+            voice = texttospeech.types.VoiceSelectionParams(language_code="en-US")
+
+            # Select the type of audio file you want returned
+            audio_config = texttospeech.types.AudioConfig(
+                audio_encoding=texttospeech.enums.AudioEncoding.LINEAR16
+            )
+
+            # Perform the text-to-speech request on the text input with the selected
+            # voice parameters and audio file type
+            def gcp_synthesize(speech_text):
+                synthesis_input = texttospeech.types.SynthesisInput(text=speech_text)
+                response = client.synthesize_speech(
+                    synthesis_input, voice, audio_config
+                )
+                return response.audio_content
+
+            self.synth_speech = gcp_synthesize

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

@@ -3,9 +3,9 @@
 
 import numpy as np
 import torch
-import pyaudio
 import klepto
 import argparse
+import warnings
 from pathlib import Path
 from .model import Tacotron2
 from glow import WaveGlow
@@ -15,9 +15,9 @@ from .text import text_to_sequence
 from .denoiser import Denoiser
 from .audio_processing import griffin_lim, postprocess_audio
 
-TTS_SAMPLE_RATE = 22050
 OUTPUT_SAMPLE_RATE = 22050
-# OUTPUT_SAMPLE_RATE = 16000
+GL_ITERS = 30
+VOCODER_WAVEGLOW, VOCODER_GL = "wavglow", "gl"
 
 # config from
 # https://github.com/NVIDIA/waveglow/blob/master/config.json
@@ -37,7 +37,7 @@ class TTSModel(object):
     def __init__(self, tacotron2_path, waveglow_path, **kwargs):
         super(TTSModel, self).__init__()
         hparams = HParams(**kwargs)
-        hparams.sampling_rate = TTS_SAMPLE_RATE
+        self.hparams = hparams
         self.model = Tacotron2(hparams)
         if torch.cuda.is_available():
             self.model.load_state_dict(torch.load(tacotron2_path)["state_dict"])
@@ -74,11 +74,11 @@ class TTSModel(object):
                 self.waveglow, n_mel_channels=hparams.n_mel_channels
             )
             self.synth_speech = klepto.safe.inf_cache(cache=self.k_cache)(
-                self.synth_speech
+                self._synth_speech
             )
         else:
             self.synth_speech = klepto.safe.inf_cache(cache=self.k_cache)(
-                self.synth_speech_gl
+                self._synth_speech_fast
             )
         self.taco_stft = TacotronSTFT(
             hparams.filter_length,
@@ -89,7 +89,7 @@ class TTSModel(object):
             mel_fmax=4000,
         )
 
-    def generate_mel_postnet(self, text):
+    def _generate_mel_postnet(self, text):
         sequence = np.array(text_to_sequence(text, ["english_cleaners"]))[None, :]
         if torch.cuda.is_available():
             sequence = torch.autograd.Variable(torch.from_numpy(sequence)).cuda().long()
@@ -101,45 +101,66 @@ class TTSModel(object):
             )
         return mel_outputs_postnet
 
-    def synth_speech_array(self, text):
-        mel_outputs_postnet = self.generate_mel_postnet(text)
+    def synth_speech_array(self, text, vocoder):
+        mel_outputs_postnet = self._generate_mel_postnet(text)
 
+        if vocoder == VOCODER_WAVEGLOW:
             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()
-        return audio
-
-    def synth_speech(self, text):
-        audio = self.synth_speech_array(text)
-
-        return postprocess_audio(
-            audio, src_rate=TTS_SAMPLE_RATE, dst_rate=OUTPUT_SAMPLE_RATE
-        )
-
-    def synth_speech_gl(self, text, griffin_iters=60):
-        mel_outputs_postnet = self.generate_mel_postnet(text)
-
+            audio = audio_t[0].data
+        elif vocoder == VOCODER_GL:
             mel_decompress = self.taco_stft.spectral_de_normalize(mel_outputs_postnet)
             mel_decompress = mel_decompress.transpose(1, 2).data.cpu()
             spec_from_mel_scaling = 1000
             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,
-            griffin_iters,
+                GL_ITERS,
             )
             audio = audio.squeeze()
+        else:
+            raise ValueError("vocoder arg should be one of [wavglow|gl]")
         audio = audio.cpu().numpy()
+        return audio
+
+    def _synth_speech(
+        self, text, speed: float = 1.0, sample_rate: int = OUTPUT_SAMPLE_RATE
+    ):
+        audio = self.synth_speech_array(text, VOCODER_WAVEGLOW)
 
         return postprocess_audio(
-            audio, tempo=0.6, src_rate=TTS_SAMPLE_RATE, dst_rate=OUTPUT_SAMPLE_RATE
+            audio,
+            src_rate=self.hparams.sampling_rate,
+            dst_rate=sample_rate,
+            tempo=speed,
+        )
+
+    def _synth_speech_fast(
+        self, text, speed: float = 1.0, sample_rate: int = OUTPUT_SAMPLE_RATE
+    ):
+        audio = self.synth_speech_array(text, VOCODER_GL)
+
+        return postprocess_audio(
+            audio,
+            tempo=speed,
+            src_rate=self.hparams.sampling_rate,
+            dst_rate=sample_rate,
         )
 
 
 def player_gen():
+    try:
+        import pyaudio
+    except ModuleNotFoundError:
+        warnings.warn("module 'pyaudio' is not installed requried for playback")
+        return
     audio_interface = pyaudio.PyAudio()
     _audio_stream = audio_interface.open(
         format=pyaudio.paInt16, channels=1, rate=OUTPUT_SAMPLE_RATE, output=True