Export multispeaker onnx (coqui-ai#2743)

TTS/tts/layers/bark/hubert/kmeans_hubert.py

@@ -16,8 +16,6 @@
 from torchaudio.functional import resample
 from transformers import HubertModel
 
-logging.root.setLevel(logging.ERROR)
-
 
 def round_down_nearest_multiple(num, divisor):
     return num // divisor * divisor

TTS/tts/models/vits.py

@@ -1876,19 +1876,20 @@ def onnx_inference(text, text_lengths, scales, sid=None):
     def load_onnx(self, model_path: str, cuda=False):
         import onnxruntime as ort
 
-        providers = ["CPUExecutionProvider" if cuda is False else "CUDAExecutionProvider"]
+        providers = [
+            "CPUExecutionProvider"
+            if cuda is False
+            else ("CUDAExecutionProvider", {"cudnn_conv_algo_search": "DEFAULT"})
+        ]
         sess_options = ort.SessionOptions()
         self.onnx_sess = ort.InferenceSession(
             model_path,
             sess_options=sess_options,
             providers=providers,
         )
 
-    def inference_onnx(self, x, x_lengths=None):
-        """ONNX inference (only single speaker models are supported)
-
-        TODO: implement multi speaker support.
-        """
+    def inference_onnx(self, x, x_lengths=None, speaker_id=None):
+        """ONNX inference"""
 
         if isinstance(x, torch.Tensor):
             x = x.cpu().numpy()
@@ -1908,7 +1909,7 @@ def inference_onnx(self, x, x_lengths=None):
                 "input": x,
                 "input_lengths": x_lengths,
                 "scales": scales,
-                "sid": None,
+                "sid": torch.tensor([speaker_id]).cpu().numpy(),
             },
         )
         return audio[0][0]