Merge pull request #16 from espnet/kan-bayashi-patch-1

Add new multi-speaker pretrained models

.github/workflows/model_test.yaml

@@ -1,6 +1,12 @@
 name: Model test
 
-on: [push, pull_request]
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    branches:
+      - master
 
 jobs:
   check_skip:

.github/workflows/unittest.yaml

@@ -1,6 +1,12 @@
 name: Unitest
 
-on: [push, pull_request]
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    branches:
+      - master
 
 jobs:
   check_skip:

ci/test_model.py

@@ -1,3 +1,6 @@
+import os
+import shutil
+
 import numpy as np
 
 from espnet2.bin.asr_inference import Speech2Text
@@ -6,37 +9,43 @@
 
 
 def _asr(model_name):
-    d = ModelDownloader()
-    speech2text = Speech2Text(**d.download_and_unpack(model_name))
+    d = ModelDownloader("downloads")
+    speech2text = Speech2Text(**d.download_and_unpack(model_name, quiet=True))
     speech = np.zeros((10000,), dtype=np.float32)
     nbests = speech2text(speech)
     text, *_ = nbests[0]
     assert isinstance(text, str)
 
 
 def _tts(model_name):
-    d = ModelDownloader()
-    text2speech = Text2Speech(**d.download_and_unpack(model_name))
-    speech = np.zeros((10000,), dtype=np.float32)
+    d = ModelDownloader("downloads")
+    text2speech = Text2Speech(**d.download_and_unpack(model_name, quiet=True))
+    inputs = {"text": "foo"}
     if text2speech.use_speech:
-        text2speech("foo", speech=speech)
-    else:
-        text2speech("foo")
+        inputs["speech"] = np.zeros((10000,), dtype=np.float32)
+    if text2speech.tts.spk_embed_dim is not None:
+        inputs["spembs"] = np.zeros((text2speech.tts.spk_embed_dim,), dtype=np.float32)
+    text2speech(**inputs)
 
 
 def test_model():
     d = ModelDownloader()
     tasks = ["asr", "tts"]
 
     for task in tasks:
-        for model_name in d.query(task=task):
-            if d.query("valid", name=model_name)[0] == "false":
-                continue
-            print(f"#### Test {model_name} ####")
-
-            if task == "asr":
-                _asr(model_name)
-            elif task == "tts":
-                _tts(model_name)
-            else:
-                raise NotImplementedError(f"task={task}")
+        for corpus in list(set(d.query("corpus", task=task))):
+            for model_name in d.query(task=task, corpus=corpus):
+                if d.query("valid", name=model_name)[0] == "false":
+                    continue
+                print(f"#### Test {model_name} ####")
+
+                if task == "asr":
+                    _asr(model_name)
+                elif task == "tts":
+                    _tts(model_name)
+                else:
+                    raise NotImplementedError(f"task={task}")
+
+            # NOTE(kan-bayashi): remove and recreate cache dir to reduce the disk usage.
+            shutil.rmtree("downloads")
+            os.makedirs("downloads")

espnet_model_zoo/downloader.py

@@ -46,7 +46,7 @@ def str_to_hash(string: Union[str, Path]) -> str:
     return hashlib.md5(str(string).encode("utf-8")).hexdigest()
 
 
-def download(url, output_path, retry: int = 3, chunk_size: int = 8192):
+def download(url, output_path, retry: int = 3, chunk_size: int = 8192, quiet=False):
     # Set retry
     session = requests.Session()
     session.mount("http://", requests.adapters.HTTPAdapter(max_retries=retry))
@@ -62,17 +62,22 @@ def download(url, output_path, retry: int = 3, chunk_size: int = 8192):
     # Write in temporary file
     with tempfile.TemporaryDirectory() as d:
         with (Path(d) / "tmp").open("wb") as f:
-            with tqdm(
-                desc=url,
-                total=file_size,
-                unit="B",
-                unit_scale=True,
-                unit_divisor=1024,
-            ) as pbar:
+            if quiet:
                 for chunk in response.iter_content(chunk_size=chunk_size):
                     if chunk:
                         f.write(chunk)
-                        pbar.update(len(chunk))
+            else:
+                with tqdm(
+                    desc=url,
+                    total=file_size,
+                    unit="B",
+                    unit_scale=True,
+                    unit_divisor=1024,
+                ) as pbar:
+                    for chunk in response.iter_content(chunk_size=chunk_size):
+                        if chunk:
+                            f.write(chunk)
+                            pbar.update(len(chunk))
 
         Path(output_path).parent.mkdir(parents=True, exist_ok=True)
         shutil.move(Path(d) / "tmp", output_path)
@@ -224,7 +229,9 @@ def unpack_local_file(self, name: str = None) -> Dict[str, Union[str, List[str]]
         # Extract files from archived file
         return unpack(filename, outdir)
 
-    def download(self, name: str = None, version: int = -1, **kwargs: str) -> str:
+    def download(
+        self, name: str = None, version: int = -1, quiet: bool = False, **kwargs: str
+    ) -> str:
         url = self.get_url(name=name, version=version, **kwargs)
         if not is_url(url) and Path(url).exists():
             return url
@@ -233,7 +240,7 @@ def download(self, name: str = None, version: int = -1, **kwargs: str) -> str:
         filename = self._get_file_name(url)
         # Download the model file if not existing
         if not (outdir / filename).exists():
-            download(url, outdir / filename)
+            download(url, outdir / filename, quiet=quiet)
 
             # Write the url for debugging
             with (outdir / "url").open("w", encoding="utf-8") as f:
@@ -261,7 +268,7 @@ def download(self, name: str = None, version: int = -1, **kwargs: str) -> str:
         return str(outdir / filename)
 
     def download_and_unpack(
-        self, name: str = None, version: int = -1, **kwargs: str
+        self, name: str = None, version: int = -1, quiet: bool = False, **kwargs: str
     ) -> Dict[str, Union[str, List[str]]]:
         url = self.get_url(name=name, version=version, **kwargs)
         if not is_url(url) and Path(url).exists():
@@ -278,7 +285,7 @@ def download_and_unpack(
                 return info
 
         # Download the file to an unique path
-        filename = self.download(url)
+        filename = self.download(url, quiet=quiet)
 
         # Extract files from archived file
         return unpack(filename, outdir)

espnet_model_zoo/table.csv

@@ -58,3 +58,23 @@ jsut,tts,kan-bayashi/jsut_tts_train_fastspeech2_transformer_teacher_raw_phn_jaco
 jsut,tts,kan-bayashi/jsut_tts_train_conformer_fastspeech2_transformer_teacher_raw_phn_jaconv_pyopenjtalk_accent_train.loss.ave,https://zenodo.org/record/4391409/files/tts_train_conformer_fastspeech2_transformer_teacher_raw_phn_jaconv_pyopenjtalk_accent_train.loss.ave.zip?download=1,24000,jp,female,1.5.1,0.9.6,acd6957,true
 chime4,asr,kamo-naoyuki/chime4_asr_train_asr_transformer3_raw_en_char_sp_valid.acc.ave,https://zenodo.org/record/4414883/files/asr_train_asr_transformer3_raw_en_char_sp_valid.acc.ave.zip?download=1,16000,en,,1.4.0,0.9.6,d5ddd5e,true
 dirha_wsj,asr,kamo-naoyuki/dirha_wsj_asr_train_asr_transformer_cmvn_raw_char_rir_scpdatadirha_irwav.scp_noise_db_range10_17_noise_scpdatadirha_noisewav.scp_speech_volume_normalize1.0_num_workers2_rir_apply_prob1._sp_valid.acc.ave,https://zenodo.org/record/4415021/files/asr_train_asr_transformer_cmvn_raw_char_rir_scpdatadirha_irwav.scp_noise_db_range10_17_noise_scpdatadirha_noisewav.scp_speech_volume_normalize1.0_num_workers2_rir_apply_prob1._sp_valid.acc.ave.zip?download=1,16000,en,,1.5.1,0.9.6,c30ce88,true
+vctk,tts,kan-bayashi/vctk_tts_train_xvector_transformer_raw_phn_tacotron_g2p_en_no_space_train.loss.ave,https://zenodo.org/record/4393279/files/tts_train_xvector_transformer_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,96ce09,true
+vctk,tts,kan-bayashi/vctk_tts_train_gst+xvector_transformer_raw_phn_tacotron_g2p_en_no_space_train.loss.ave,https://zenodo.org/record/4393277/files/tts_train_gst%2Bxvector_transformer_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,96ce09,true
+vctk,tts,kan-bayashi/vctk_tts_train_xvector_tacotron2_raw_phn_tacotron_g2p_en_no_space_train.loss.ave,https://zenodo.org/record/4394600/files/tts_train_xvector_tacotron2_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,96ce09,true
+vctk,tts,kan-bayashi/vctk_tts_train_gst+xvector_tacotron2_raw_phn_tacotron_g2p_en_no_space_train.loss.ave,https://zenodo.org/record/4394598/files/tts_train_gst%2Bxvector_tacotron2_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,96ce09,true
+vctk,tts,kan-bayashi/vctk_tts_train_xvector_conformer_fastspeech2_transformer_teacher_raw_phn_tacotron_g2p_en_no_space_train.loss.ave,https://zenodo.org/record/4394602/files/tts_train_xvector_conformer_fastspeech2_transformer_teacher_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,96ce09,true
+vctk,tts,kan-bayashi/vctk_tts_train_gst+xvector_conformer_fastspeech2_transformer_teacher_raw_phn_tacotron_g2p_en_no_space_train.loss.ave,https://zenodo.org/record/4394608/files/tts_train_gst%2Bxvector_conformer_fastspeech2_transformer_teacher_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,96ce09,true
+libritts,tts,kan-bayashi/libritts_tts_train_xvector_trasnformer_raw_phn_tacotron_g2p_en_no_space_train.loss.ave,https://zenodo.org/record/4409704/files/tts_train_xvector_transformer_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,861431,true
+libritts,tts,kan-bayashi/libritts_tts_train_gst+xvector_trasnformer_raw_phn_tacotron_g2p_en_no_space_train.loss.ave,https://zenodo.org/record/4409702/files/tts_train_gst%2Bxvector_transformer_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,861431,true
+libritts,tts,kan-bayashi/libritts_tts_train_xvector_conformer_fastspeech2_transformer_teacher_raw_phn_tacotron_g2p_en_no_space_train.loss,https://zenodo.org/record/4418754/files/tts_train_xvector_conformer_fastspeech2_transformer_teacher_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,861431,true
+libritts,tts,kan-bayashi/libritts_tts_train_gst+xvector_conformer_fastspeech2_transformer_teacher_raw_phn_tacotron_g2p_en_no_space_train.loss,https://zenodo.org/record/4418774/files/tts_train_gst%2Bxvector_conformer_fastspeech2_transformer_teacher_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,861431,true
+vctk,tts,kan-bayashi/vctk_xvector_transformer,https://zenodo.org/record/4393279/files/tts_train_xvector_transformer_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,96ce09,true
+vctk,tts,kan-bayashi/vctk_gst+xvector_transformer,https://zenodo.org/record/4393277/files/tts_train_gst%2Bxvector_transformer_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,96ce09,true
+vctk,tts,kan-bayashi/vctk_xvector_tacotron2,https://zenodo.org/record/4394600/files/tts_train_xvector_tacotron2_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,96ce09,true
+vctk,tts,kan-bayashi/vctk_gst+xvector_tacotron2,https://zenodo.org/record/4394598/files/tts_train_gst%2Bxvector_tacotron2_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,96ce09,true
+vctk,tts,kan-bayashi/vctk_xvector_conformer_fastspeech2,https://zenodo.org/record/4394602/files/tts_train_xvector_conformer_fastspeech2_transformer_teacher_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,96ce09,true
+vctk,tts,kan-bayashi/vctk_gst+xvector_conformer_fastspeech2,https://zenodo.org/record/4394608/files/tts_train_gst%2Bxvector_conformer_fastspeech2_transformer_teacher_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,96ce09,true
+libritts,tts,kan-bayashi/libritts_xvector_trasnformer,https://zenodo.org/record/4409704/files/tts_train_xvector_transformer_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,861431,true
+libritts,tts,kan-bayashi/libritts_gst+xvector_trasnformer,https://zenodo.org/record/4409702/files/tts_train_gst%2Bxvector_transformer_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,861431,true
+libritts,tts,kan-bayashi/libritts_xvector_conformer_fastspeech2,https://zenodo.org/record/4418754/files/tts_train_xvector_conformer_fastspeech2_transformer_teacher_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,861431,true
+libritts,tts,kan-bayashi/libritts_gst+xvector_conformer_fastspeech2,https://zenodo.org/record/4418774/files/tts_train_gst%2Bxvector_conformer_fastspeech2_transformer_teacher_raw_phn_tacotron_g2p_en_no_space_train.loss.ave.zip?download=1,24000,en,,1.5.1,0.9.6,861431,true

setup.py

@@ -29,7 +29,7 @@
 dirname = os.path.dirname(__file__)
 setup(
     name="espnet_model_zoo",
-    version="0.0.0a21",
+    version="0.0.0a22",
     url="http://github.com/espnet/espnet_model_zoo",
     description="ESPnet Model Zoo",
     long_description=open(os.path.join(dirname, "README.md"), encoding="utf-8").read(),