init

experiment_voxceleb1.py

@@ -0,0 +1,43 @@
+import json
+import os
+from os.path import join as p_join
+from tqdm import tqdm
+
+from datasets import load_dataset
+
+from model_meta_voice import MetaVoiceSE
+from model_pyannote_embedding import PyannoteSE
+from model_w2v_bert import W2VBertSE
+
+
+cache_dir = p_join("experiment_cache", "voxceleb1")
+voxceleb1_dataset = load_dataset("asahi417/voxceleb1-test-split", split="test")
+print(f"voxceleb1_dataset: {len(voxceleb1_dataset)}")
+
+
+def get_embedding(model_class, model_name: str):
+    file_path = p_join(cache_dir, f"embedding.{model_name}.json")
+    if os.path.exists(file_path):
+        return
+    model = model_class()
+    embeddings = []
+    for i in tqdm(voxceleb1_dataset, total=len(voxceleb1_dataset)):
+        v = model.get_speaker_embedding(i["audio"]["array"], i["audio"]["sampling_rate"])
+        embeddings.append({
+            "model": model_name,
+            "embedding": v.tolist(),
+            "sampling_rate": i["audio"]["sampling_rate"],
+            "id": i["id"],
+            "speaker_id": i["speaker_id"]
+        })
+    with open(file_path, "w") as f:
+        f.write("\n".join([json.dumps(i) for i in embeddings]))
+
+
+if __name__ == '__main__':
+    # cache embedding
+    get_embedding(MetaVoiceSE, "meta_voice_se")
+    get_embedding(PyannoteSE, "pyannote_se")
+    get_embedding(W2VBertSE, "w2v_bert_se")
+
+