KALDI之aishell之V1模型續4(最終的結果)
阿新 • • 發佈:2018-11-14
成功結束了aishell之V1模型:
eer=0.049%
sid/extract_ivectors.sh --cmd run.pl --mem 4G --nj 10 exp/extractor_male data/test/enroll exp/ivector_enroll_1024 sid/extract_ivectors.sh: extracting iVectors sid/extract_ivectors.sh: combining iVectors across jobs sid/extract_ivectors.sh: computing mean of iVectors for each speaker and length-normalizing sid/extract_ivectors.sh --cmd run.pl --mem 4G --nj 10 exp/extractor_male data/test/eval exp/ivector_eval_1024 sid/extract_ivectors.sh: extracting iVectors sid/extract_ivectors.sh: combining iVectors across jobs sid/extract_ivectors.sh: computing mean of iVectors for each speaker and length-normalizing compute-eer - LOG (compute-eer[5.5.39~1-88f23]:main():compute-eer.cc:136) Equal error rate is 0.0487076%, at threshold -12.62450.04871
錯誤率太低,懷疑自己拿了訓練資料當測試資料 ,然後尷尬成訓練自己又識別自己嗎
data下的feats.scp
test下的wav.scp
繼續檢視打分部分總指令碼:
找run.sh中trials這個第一次出現
根據trails路徑開啟檔案
把data/test的utt2spk開啟看看
github原始碼中的錯誤率Equal error rate is 0.140528%, at threshold -12.018;
我的錯誤率 Equal error rate is 0.0487076%, at threshold -12.6245