feat: 增加情感检测与标点识别

utils/stt/speaker_ver_utils.py

@@ -0,0 +1,75 @@
+from modelscope.pipelines import pipeline
+import numpy as np
+import os
+import pdb
+ERES2NETV2 = {
+    "task": 'speaker-verification',
+    "model_name": 'damo/speech_eres2netv2_sv_zh-cn_16k-common',
+    "model_revision": 'v1.0.1',
+    "save_embeddings": False
+}
+
+# 保存 embedding 的路径
+DEFALUT_SAVE_PATH = r".\takway\savePath"
+
+class speaker_verfication:
+    def __init__(self, 
+                task='speaker-verification',
+                model_name='damo/speech_eres2netv2_sv_zh-cn_16k-common',
+                model_revision='v1.0.1',
+                device="cuda",
+                save_embeddings=False):
+        self.pipeline = pipeline(
+                task=task,
+                model=model_name,
+                model_revision=model_revision,
+                device=device)
+        self.save_embeddings = save_embeddings
+
+    def wav2embeddings(self, speaker_1_wav):
+        result = self.pipeline([speaker_1_wav], output_emb=True)
+        speaker_1_emb = result['embs'][0]
+        return speaker_1_emb
+
+    def _verifaction(self, speaker_1_wav, speaker_2_wav, threshold, save_path):
+        if not self.save_embeddings:
+            result = self.pipeline([speaker_1_wav, speaker_2_wav], thr=threshold)
+            return result["text"]
+        else:
+            result = self.pipeline([speaker_1_wav, speaker_2_wav], thr=threshold, output_emb=True)
+            speaker1_emb = result["embs"][0]
+            speaker2_emb = result["embs"][1]
+            np.save(os.path.join(save_path, "speaker_1.npy"), speaker1_emb)
+            return result['outputs']["text"]
+        
+    def _verifaction_from_embedding(self, base_emb, speaker_2_wav, threshold):
+        base_emb = np.load(base_emb)
+        result = self.pipeline([speaker_2_wav], output_emb=True)
+        speaker2_emb = result["embs"][0]
+        similarity = np.dot(base_emb, speaker2_emb) / (np.linalg.norm(base_emb) * np.linalg.norm(speaker2_emb))
+        if similarity > threshold:
+            return "yes"
+        else:
+            return "no"
+    
+    def verfication(self, 
+                    base_emb, 
+                    speaker_emb, 
+                    threshold=0.333, ):
+        return np.dot(base_emb, speaker_emb) / (np.linalg.norm(base_emb) * np.linalg.norm(speaker_emb)) > threshold
+        
+if __name__ == '__main__':
+    verifier = speaker_verfication(**ERES2NETV2)
+
+    verifier = speaker_verfication(save_embeddings=False)
+    result = verifier.verfication(base_emb=None, speaker_1_wav=r"C:\Users\bing\Downloads\speaker1_a_cn_16k.wav", 
+                                  speaker_2_wav=r"C:\Users\bing\Downloads\speaker2_a_cn_16k.wav",
+                                  threshold=0.333,
+                                  save_path=r"D:\python\irving\takway_base-main\savePath"
+                                  )
+    print("---")
+    print(result)
+    print(verifier.verfication(r"D:\python\irving\takway_base-main\savePath\speaker_1.npy",
+                               speaker_2_wav=r"C:\Users\bing\Downloads\speaker1_b_cn_16k.wav",
+                               threshold=0.333,
+                               ))