提交

kongdeqiang

5 天以前 28ccfbfc51068a663a80764e14074df5edf2b5ba

 tests/test_sv_inference_pipeline.py

@@ -1,5 +1,6 @@
import unittest

import numpy as np
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
from modelscope.utils.logger import get_logger
@@ -11,37 +12,37 @@
    def test_funasr_path(self):
        import funasr
        import os

        logger.info("run_dir:{0} ; funasr_path: {1}".format(os.getcwd(), funasr.__file__))

    def test_inference_pipeline(self):
        inference_sv_pipline = pipeline(
            task=Tasks.speaker_verification,
            model='damo/speech_xvector_sv-zh-cn-cnceleb-16k-spk3465-pytorch'
            model="damo/speech_xvector_sv-zh-cn-cnceleb-16k-spk3465-pytorch",
        )
        # 提取不同句子的说话人嵌入码

        # the same speaker
        rec_result = inference_sv_pipline(
            audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/sv_example_enroll.wav')
        enroll = rec_result["spk_embedding"]
            audio_in=(
                "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/sv_example_enroll.wav",
                "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/sv_example_same.wav",
            )
        )
        assert (
            abs(rec_result["scores"][0] - 0.85) < 0.1 and abs(rec_result["scores"][1] - 0.14) < 0.1
        )
        logger.info(f"Similarity {rec_result['scores']}")

        # different speaker
        rec_result = inference_sv_pipline(
            audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/sv_example_same.wav')
        same = rec_result["spk_embedding"]

        rec_result = inference_sv_pipline(
            audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/sv_example_different.wav')
        different = rec_result["spk_embedding"]

        # 对相同的说话人计算余弦相似度
        sv_threshold = 0.9465
        same_cos = np.sum(enroll * same) / (np.linalg.norm(enroll) * np.linalg.norm(same))
        same_cos = max(same_cos - sv_threshold, 0.0) / (1.0 - sv_threshold) * 100.0
        logger.info("Similarity: {}".format(same_cos))

        # 对不同的说话人计算余弦相似度
        diff_cos = np.sum(enroll * different) / (np.linalg.norm(enroll) * np.linalg.norm(different))
        diff_cos = max(diff_cos - sv_threshold, 0.0) / (1.0 - sv_threshold) * 100.0
        logger.info("Similarity: {}".format(diff_cos))
            audio_in=(
                "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/sv_example_enroll.wav",
                "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/sv_example_different.wav",
            )
        )
        assert abs(rec_result["scores"][0] - 0.0) < 0.1 and abs(rec_result["scores"][1] - 1.0) < 0.1
        logger.info(f"Similarity {rec_result['scores']}")


if __name__ == '__main__':
if __name__ == "__main__":
    unittest.main()