init

experiment_speaker_verification.py

@@ -18,6 +18,7 @@ from model_meta_voice import MetaVoiceSE
 from model_pyannote_embedding import PyannoteSE
 from model_w2v_bert import W2VBertSE
 from model_clap import ClapSE, ClapGeneralSE
+from model_xls import XLSRSE
 
 
 def get_embedding(model_class, model_name: str, dataset_name: str, data_split: str):
@@ -97,9 +98,9 @@ def cluster_embedding(model_name, dataset_name, label_name: str):
     plt.gca().set_aspect('equal', 'datalim')
     plt.legend(handles=scatter.legend_elements(num=len(label_type))[0],
                labels=label_type,
-               bbox_to_anchor=(1.05, 1),
+               bbox_to_anchor=(1.04, 1),
                borderaxespad=0,
-               loc='lower left',
+               loc='upper left',
                ncol=3 if len(label2id) > 12 else 1)
     plt.savefig(figure_path, bbox_inches='tight', dpi=600)
 
@@ -115,35 +116,40 @@ def analyze_embedding(model_name: str, dataset_name: str, n_shot: int = 5, n_cro
 
 
 if __name__ == '__main__':
-    get_embedding(MetaVoiceSE, "meta_voice_se", "asahi417/voxceleb1-test-split", "test")
-    get_embedding(PyannoteSE, "pyannote_se", "asahi417/voxceleb1-test-split", "test")
-    get_embedding(W2VBertSE, "w2v_bert_se", "asahi417/voxceleb1-test-split", "test")
-    get_embedding(ClapSE, "clap_se", "asahi417/voxceleb1-test-split", "test")
-    get_embedding(ClapGeneralSE, "clap_general_se", "asahi417/voxceleb1-test-split", "test")
-
-    get_embedding(MetaVoiceSE, "meta_voice_se", "ylacombe/expresso", "train")
-    get_embedding(PyannoteSE, "pyannote_se", "ylacombe/expresso", "train")
-    get_embedding(W2VBertSE, "w2v_bert_se", "ylacombe/expresso", "train")
-    get_embedding(ClapSE, "clap_se", "ylacombe/expresso", "train")
-    get_embedding(ClapGeneralSE, "clap_general_se", "ylacombe/expresso", "train")
-
-    cluster_embedding("meta_voice_se", "asahi417/voxceleb1-test-split", "speaker_id")
-    cluster_embedding("pyannote_se", "asahi417/voxceleb1-test-split", "speaker_id")
-    cluster_embedding("w2v_bert_se", "asahi417/voxceleb1-test-split", "speaker_id")
-    cluster_embedding("clap_se", "asahi417/voxceleb1-test-split", "speaker_id")
-    cluster_embedding("clap_general_se", "asahi417/voxceleb1-test-split", "speaker_id")
-
-    cluster_embedding("meta_voice_se", "ylacombe/expresso", "speaker_id")
-    cluster_embedding("pyannote_se", "ylacombe/expresso", "speaker_id")
-    cluster_embedding("w2v_bert_se", "ylacombe/expresso", "speaker_id")
-    cluster_embedding("clap_se", "ylacombe/expresso", "speaker_id")
-    cluster_embedding("clap_general_se", "ylacombe/expresso", "speaker_id")
-
-    cluster_embedding("meta_voice_se", "ylacombe/expresso", "style")
-    cluster_embedding("pyannote_se", "ylacombe/expresso", "style")
-    cluster_embedding("w2v_bert_se", "ylacombe/expresso", "style")
-    cluster_embedding("clap_se", "ylacombe/expresso", "style")
-    cluster_embedding("clap_general_se", "ylacombe/expresso", "style")
+    # get_embedding(MetaVoiceSE, "meta_voice_se", "asahi417/voxceleb1-test-split", "test")
+    # get_embedding(PyannoteSE, "pyannote_se", "asahi417/voxceleb1-test-split", "test")
+    # get_embedding(W2VBertSE, "w2v_bert_se", "asahi417/voxceleb1-test-split", "test")
+    # get_embedding(ClapSE, "clap_se", "asahi417/voxceleb1-test-split", "test")
+    # get_embedding(ClapGeneralSE, "clap_general_se", "asahi417/voxceleb1-test-split", "test")
+    get_embedding(XLSRSE, "xlsr_se", "asahi417/voxceleb1-test-split", "test")
+
+    # get_embedding(MetaVoiceSE, "meta_voice_se", "ylacombe/expresso", "train")
+    # get_embedding(PyannoteSE, "pyannote_se", "ylacombe/expresso", "train")
+    # get_embedding(W2VBertSE, "w2v_bert_se", "ylacombe/expresso", "train")
+    # get_embedding(ClapSE, "clap_se", "ylacombe/expresso", "train")
+    # get_embedding(ClapGeneralSE, "clap_general_se", "ylacombe/expresso", "train")
+    get_embedding(XLSRSE, "xlsr_se", "ylacombe/expresso", "train")
+
+    # cluster_embedding("meta_voice_se", "asahi417/voxceleb1-test-split", "speaker_id")
+    # cluster_embedding("pyannote_se", "asahi417/voxceleb1-test-split", "speaker_id")
+    # cluster_embedding("w2v_bert_se", "asahi417/voxceleb1-test-split", "speaker_id")
+    # cluster_embedding("clap_se", "asahi417/voxceleb1-test-split", "speaker_id")
+    # cluster_embedding("clap_general_se", "asahi417/voxceleb1-test-split", "speaker_id")
+    cluster_embedding("xlsr_se", "asahi417/voxceleb1-test-split", "speaker_id")
+    #
+    # cluster_embedding("meta_voice_se", "ylacombe/expresso", "speaker_id")
+    # cluster_embedding("pyannote_se", "ylacombe/expresso", "speaker_id")
+    # cluster_embedding("w2v_bert_se", "ylacombe/expresso", "speaker_id")
+    # cluster_embedding("clap_se", "ylacombe/expresso", "speaker_id")
+    # cluster_embedding("clap_general_se", "ylacombe/expresso", "speaker_id")
+    cluster_embedding("xlsr_se", "ylacombe/expresso", "speaker_id")
+    #
+    # cluster_embedding("meta_voice_se", "ylacombe/expresso", "style")
+    # cluster_embedding("pyannote_se", "ylacombe/expresso", "style")
+    # cluster_embedding("w2v_bert_se", "ylacombe/expresso", "style")
+    # cluster_embedding("clap_se", "ylacombe/expresso", "style")
+    # cluster_embedding("clap_general_se", "ylacombe/expresso", "style")
+    cluster_embedding("xlsr_se", "ylacombe/expresso", "style")
 
 
 

model_xls.py

@@ -0,0 +1,29 @@
+"""Meta's XLS-R based speaker embedding.
+- feature dimension: 768
+- source: https://huggingface.co/facebook/wav2vec2-large-xlsr-53
+"""
+from typing import Optional
+
+import torch
+import librosa
+import numpy as np
+from transformers import AutoFeatureExtractor, AutoModelForPreTraining
+
+
+class XLSRSE:
+    def __init__(self):
+        self.processor = AutoFeatureExtractor.from_pretrained("facebook/wav2vec2-large-xlsr-53")
+        self.model = AutoModelForPreTraining.from_pretrained("facebook/wav2vec2-large-xlsr-53")
+        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.projected_states.mean(1).cpu().numpy()[0]
+        # return outputs.projected_quantized_states.mean(1).cpu().numpy()[0]

test.py

@@ -3,26 +3,34 @@ from model_clap import ClapSE
 from model_meta_voice import MetaVoiceSE
 from model_pyannote_embedding import PyannoteSE
 from model_w2v_bert import W2VBertSE
+from model_xls import XLSRSE
 
 
 def test():
     wav, sr = librosa.load("sample.wav")
+    print("XLS-R")
+    model = XLSRSE()
+    v = model.get_speaker_embedding(wav, sr)
+    print(v.shape)
+    model = ClapSE()
+    v = model.get_speaker_embedding(wav, sr)
+    print(v.shape)
     print("CLAP")
     model = ClapSE()
     v = model.get_speaker_embedding(wav, sr)
     print(v.shape)
-    # print("MetaVoiceSE")
-    # model = MetaVoiceSE()
-    # v = model.get_speaker_embedding(wav, sr)
-    # print(v.shape)
-    # print("PyannoteSE")
-    # model = PyannoteSE()
-    # v = model.get_speaker_embedding(wav, sr)
-    # print(v.shape)
-    # print("W2VBertSE")
-    # model = W2VBertSE()
-    # v = model.get_speaker_embedding(wav, sr)
-    # print(v.shape)
+    print("MetaVoiceSE")
+    model = MetaVoiceSE()
+    v = model.get_speaker_embedding(wav, sr)
+    print(v.shape)
+    print("PyannoteSE")
+    model = PyannoteSE()
+    v = model.get_speaker_embedding(wav, sr)
+    print(v.shape)
+    print("W2VBertSE")
+    model = W2VBertSE()
+    v = model.get_speaker_embedding(wav, sr)
+    print(v.shape)
 
 
 if __name__ == '__main__':