init

attach_speaker_embedding_s2s.py

@@ -20,6 +20,15 @@ if se_model == "metavoice":
 elif se_model == "pyannote":
     from speaker_embedding_pyannote import PyannoteSE
     speaker_embedder = PyannoteSE()
+elif se_model == "w2vbert-600m":
+    from speaker_embedding_hf import Wav2VecEmbedding
+    speaker_embedder = Wav2VecEmbedding()
+elif se_model == "xlsr-2b":
+    from speaker_embedding_hf import XLSR2BEmbedding
+    speaker_embedder = XLSR2BEmbedding()
+elif se_model == "hubert-xl":
+    from speaker_embedding_hf import HuBERTXLEmbedding
+    speaker_embedder = HuBERTXLEmbedding()
 else:
     raise ValueError(f"unknown speaker embedding: {se_model}")
 
@@ -44,9 +53,14 @@ print(f"Num examples (after filtering): {len(dataset)}")
 
 def speaker_embedding(example):
     for side in sides:
-        example[f"{side}.audio.speaker_embedding"] = speaker_embedder.get_speaker_embedding(
+        embedding = speaker_embedder.get_speaker_embedding(
             example[f"{side}.audio"]["array"], example[f"{side}.audio"]["sampling_rate"]
         )
+        if embedding.ndim == 1:
+            example[f"{side}.audio.speaker_embedding"] = embedding
+        else:
+            example[f"{side}.audio.speaker_embedding"] = embedding.mean(0)
+            example[f"{side}.audio.speaker_embedding.full"] = embedding
     return example
 
 

speaker_embedding_clap.py

@@ -1,35 +0,0 @@
-"""CLAP embedding.
-- feature dimension: 512
-- source: https://huggingface.co/laion/larger_clap_music_and_speech
-"""
-from typing import Optional
-
-import torch
-import librosa
-import numpy as np
-from transformers import ClapModel, ClapProcessor
-
-
-class ClapSE:
-    def __init__(self, ckpt: str = "laion/larger_clap_music_and_speech"):
-        self.model = ClapModel.from_pretrained(ckpt)
-        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-        self.model.to(self.device)
-        self.model.eval()
-        self.processor = ClapProcessor.from_pretrained(ckpt)
-
-    def get_speaker_embedding(self, wav: np.ndarray, sampling_rate: Optional[int] = None) -> np.ndarray:
-        if sampling_rate != self.processor.feature_extractor.sampling_rate:
-            wav = librosa.resample(wav, orig_sr=sampling_rate, target_sr=self.processor.feature_extractor.sampling_rate)
-        inputs = self.processor(
-            audios=wav, sampling_rate=self.processor.feature_extractor.sampling_rate, return_tensors="pt"
-        )
-        with torch.no_grad():
-            outputs = self.model.get_audio_features(**{k: v.to(self.device) for k, v in inputs.items()})
-        return outputs.cpu().numpy()[0]
-
-
-class ClapGeneralSE(ClapSE):
-
-    def __init__(self):
-        super().__init__(ckpt="laion/larger_clap_general")

speaker_embedding_hf.py

@@ -0,0 +1,72 @@
+"""Meta's w2vBERT based speaker embedding."""
+from typing import Optional
+
+import torch
+import librosa
+import numpy as np
+from transformers import AutoModel, AutoFeatureExtractor
+
+
+############
+# W2V BERT #
+############
+class W2VBERTEmbedding:
+    def __init__(self, ckpt: str = "facebook/w2v-bert-2.0"):
+        self.processor = AutoFeatureExtractor.from_pretrained(ckpt)
+        self.model = AutoModel.from_pretrained(ckpt)
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        self.model.to(self.device)
+        self.model.eval()
+
+    def get_speaker_embedding(self, wav: np.ndarray, sampling_rate: Optional[int] = None) -> np.ndarray:
+        # audio file is decoded on the fly
+        if sampling_rate != self.processor.sampling_rate:
+            wav = librosa.resample(wav, orig_sr=sampling_rate, target_sr=self.processor.sampling_rate)
+        inputs = self.processor(wav, sampling_rate=self.processor.sampling_rate, return_tensors="pt")
+        with torch.no_grad():
+            outputs = self.model(**{k: v.to(self.device) for k, v in inputs.items()})
+        return outputs.last_hidden_state.cpu().numpy()[0]
+
+
+##########
+# HuBERT #
+##########
+class HuBERTXLEmbedding(W2VBERTEmbedding):
+    def __init__(self):
+        super().__init__("facebook/hubert-xlarge-ll60k")
+
+
+class HuBERTLargeEmbedding(W2VBERTEmbedding):
+    def __init__(self):
+        super().__init__("facebook/hubert-large-ll60k")
+
+
+class HuBERTBaseEmbedding(W2VBERTEmbedding):
+    def __init__(self):
+        super().__init__("facebook/hubert-base-ls960")
+
+
+###########
+# wav2vec #
+###########
+class Wav2VecEmbedding(W2VBERTEmbedding):
+    def __init__(self):
+        super().__init__("facebook/wav2vec2-large-xlsr-53")
+
+
+#########
+# XLS-R #
+#########
+class XLSR2BEmbedding(W2VBERTEmbedding):
+    def __init__(self):
+        super().__init__("facebook/wav2vec2-xls-r-2b")
+
+
+class XLSR1BEmbedding(W2VBERTEmbedding):
+    def __init__(self):
+        super().__init__("facebook/wav2vec2-xls-r-1b")
+
+
+class XLSR300MEmbedding(W2VBERTEmbedding):
+    def __init__(self):
+        super().__init__("facebook/wav2vec2-xls-r-300m")