From 16c834bcf8c5d0c891f25b307f0241c1383943de Mon Sep 17 00:00:00 2001
From: 志浩 <neo.dzh@alibaba-inc.com>
Date: 星期二, 01 八月 2023 20:53:55 +0800
Subject: [PATCH] TOLD/SOND: update finetune.sh
---
egs/callhome/diarization/sond/finetune.sh | 18 +++++++++++-------
egs/callhome/diarization/sond/script/extract_nonoverlap_segments.py | 2 +-
2 files changed, 12 insertions(+), 8 deletions(-)
diff --git a/egs/callhome/diarization/sond/finetune.sh b/egs/callhome/diarization/sond/finetune.sh
index e82dcae..6978343 100644
--- a/egs/callhome/diarization/sond/finetune.sh
+++ b/egs/callhome/diarization/sond/finetune.sh
@@ -221,13 +221,17 @@
done
fi
-
-# You will get a DER like:
-# iter0: 9.68 10.51
-# iter1:
-# iter2:
-# iter3:
-# iter4:
+# In this stage, we need the raw waveform files of Callhome corpus.
+# Due to the data license, we can't provide them, please get them additionally.
+# And convert the sph files to wav files (use scripts/dump_pipe_wav.py).
+# Then find the wav files to construct wav.scp and put it at data/callhome2/wav.scp.
+# After iteratively perform SOAP, you will get DER results like:
+# iters| oracle_vad | system_vad
+# iter_0: 9.68 | 10.51
+# iter_1: 9.26 | 10.14 (reported in the paper)
+# iter_2: 9.18 | 10.08
+# iter_3: 9.24 | 10.15
+# iter_4: 9.27 | 10.17
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
for dset in ${test_sets}; do
echo "stage 4: Evaluating finetuned system on ${dset} set with medfilter_size=83 clustering=EEND-OLA"
diff --git a/egs/callhome/diarization/sond/script/extract_nonoverlap_segments.py b/egs/callhome/diarization/sond/script/extract_nonoverlap_segments.py
index a4e43a2..37f995a 100644
--- a/egs/callhome/diarization/sond/script/extract_nonoverlap_segments.py
+++ b/egs/callhome/diarization/sond/script/extract_nonoverlap_segments.py
@@ -93,7 +93,7 @@
extracted_spk = []
count = 1
for st, ed, spk in tqdm(turns, total=len(turns), ascii=True, disable=args.no_pbar):
- if (ed - st) >= args.min_dur * args.sr and wav[st: ed] >= args.min_dur * args.sr:
+ if (ed - st) >= args.min_dur * args.sr and len(wav[st: ed]) >= args.min_dur * args.sr:
seg = wav[st: ed]
save_path = os.path.join(args.out_dir, mid, spk, "{}_U{:04d}.wav".format(spk, count))
if not os.path.exists(os.path.dirname(save_path)):
--
Gitblit v1.9.1