| | |
|---|
| | | count=1 |
|---|
| | | |
|---|
| | | # general configuration |
|---|
| | | stage=1 |
|---|
| | | stop_stage=1 |
|---|
| | | stage=0 |
|---|
| | | stop_stage=10 |
|---|
| | | # number of jobs for data process |
|---|
| | | nj=16 |
|---|
| | | sr=8000 |
|---|
| | |
|---|
| | | sv_exp_dir=exp/speech_xvector_sv-en-us-callhome-8k-spk6135-pytorch |
|---|
| | | |
|---|
| | | if [ ! -e ${sv_exp_dir} ]; then |
|---|
| | | echo "start to download sv models" |
|---|
| | | git lfs install |
|---|
| | | git clone https://www.modelscope.cn/damo/speech_xvector_sv-en-us-callhome-8k-spk6135-pytorch.git |
|---|
| | | mv speech_xvector_sv-en-us-callhome-8k-spk6135-pytorch ${expdir}/ |
|---|
| | | echo "Done." |
|---|
| | | fi |
|---|
| | | |
|---|
| | | for dset in callhome1/nonoverlap_0s callhome2/nonoverlap_0s; do |
|---|
| | | echo "Start to extract speaker embeddings for ${dset}" |
|---|
| | | key_file=${datadir}/${dset}/wav.scp |
|---|
| | | num_scp_file="$(<${key_file} wc -l)" |
|---|
| | | _nj=$([ $inference_nj -le $num_scp_file ] && echo "$inference_nj" || echo "$num_scp_file") |
|---|
| | |
|---|
| | | --sv_model_file ${sv_exp_dir}/sv.pth \ |
|---|
| | | --output_dir "${_logdir}"/output.JOB |
|---|
| | | cat ${_logdir}/output.*/xvector.scp | sort > ${datadir}/${dset}/utt2xvec |
|---|
| | | |
|---|
| | | python script/calc_num_frames.py ${key_file} ${datadir}/${dset}/utt2num_frames |
|---|
| | | echo "Done." |
|---|
| | | done |
|---|
| | | |
|---|
| | | fi |
|---|
| | |
|---|
| | | python -Wignore script/calc_real_meeting_frame_labels.py \ |
|---|
| | | ${datadir}/${dset} ${dumpdir}/${dset}/labels \ |
|---|
| | | --n_spk 8 --frame_shift 0.01 --nj 16 --sr 8000 |
|---|
| | | find `pwd`/${dumpdir}/${dset}/labels -iname "*.lbl.mat" | awk -F'[/.]' '{print $(NF-2),$0}' | sort > ${datadir}/${dset}/labels.scp |
|---|
| | | find `pwd`/${dumpdir}/${dset}/labels/ -iname "*.lbl.mat" | awk -F'[/.]' '{print $(NF-2),$0}' | sort > ${datadir}/${dset}/labels.scp |
|---|
| | | done |
|---|
| | | |
|---|
| | | fi |
|---|
| | |
|---|
| | | |
|---|
| | | echo "Stage 6: start to dump for callhome1." |
|---|
| | | python -Wignore script/dump_meeting_chunks.py --dir ${data_dir} \ |
|---|
| | | --out ${dumpdir}/callhome1/dumped_files/data --n_spk 16 --no_pbar --sr 8000 --mode test \ |
|---|
| | | --out ${dumpdir}/callhome1/dumped_files/data --n_spk 16 --no_pbar --sr 8000 --mode train \ |
|---|
| | | --chunk_size 1600 --chunk_shift 400 --add_mid_to_speaker true |
|---|
| | | |
|---|
| | | mkdir -p ${datadir}/callhome1/dumped_files |
|---|
| | |
|---|
| | | # Finetune model on callhome1, this will take about 1.5 hours. |
|---|
| | | if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then |
|---|
| | | echo "Stage 7: Finetune pretrained model on callhome1." |
|---|
| | | |
|---|
| | | if [ ! -e ${expdir}/speech_diarization_sond-en-us-swbd_sre-8k-n16k4-pytorch ]; then |
|---|
| | | echo "start to download pretrained models" |
|---|
| | | git lfs install |
|---|
| | | git clone https://www.modelscope.cn/damo/speech_diarization_sond-en-us-swbd_sre-8k-n16k4-pytorch.git |
|---|
| | | mv speech_diarization_sond-en-us-swbd_sre-8k-n16k4-pytorch ${expdir}/ |
|---|
| | | echo "Done." |
|---|
| | | fi |
|---|
| | | |
|---|
| | | world_size=$gpu_num # run on one machine |
|---|
| | | mkdir -p ${expdir}/${model_dir} |
|---|
| | | mkdir -p ${expdir}/${model_dir}/log |
|---|
| | |
|---|
| | | --valid_data_path_and_name_and_type ${datadir}/${valid_set}/dumped_files/profile.scp,profile,kaldi_ark \ |
|---|
| | | --valid_data_path_and_name_and_type ${datadir}/${valid_set}/dumped_files/label.scp,binary_labels,kaldi_ark \ |
|---|
| | | --valid_shape_file ${expdir}/${valid_set}_states/speech_shape \ |
|---|
| | | --init_param exp/pretrained_models/phase2.pth \ |
|---|
| | | --init_param ${expdir}/speech_diarization_sond-en-us-swbd_sre-8k-n16k4-pytorch/sond.pth \ |
|---|
| | | --unused_parameters true \ |
|---|
| | | ${init_opt} \ |
|---|
| | | ${freeze_opt} \ |
|---|
| | |
|---|
| | | |
|---|
| | | # Scoring for finetuned model, you may get a DER like: |
|---|
| | | # oracle_vad | system_vad |
|---|
| | | # 7.28 | 8.06 |
|---|
| | | # 7.32 | 8.14 |
|---|
| | | if [ ${stage} -le 9 ] && [ ${stop_stage} -ge 9 ]; then |
|---|
| | | echo "stage 9: Scoring finetuned models" |
|---|
| | | if [ ! -e dscore ]; then |
|---|
| | |
|---|
| | | # 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: 9.63 | 10.43 |
|---|
| | | # iter_1: 9.17 | 10.03 |
|---|
| | | # iter_2: 9.11 | 9.98 |
|---|
| | | # iter_3: 9.08 | 9.96 |
|---|
| | | # iter_4: 9.07 | 9.95 |
|---|
| | | # iter_0: 9.58 | 10.46 |
|---|
| | | # iter_1: 9.22 | 10.15 |
|---|
| | | # iter_2: 9.21 | 10.14 |
|---|
| | | # iter_3: 9.30 | 10.24 |
|---|
| | | # iter_4: 9.29 | 10.23 |
|---|
| | | if [ ${stage} -le 10 ] && [ ${stop_stage} -ge 10 ]; then |
|---|
| | | if [ ! -e ${expdir}/speech_xvector_sv-en-us-callhome-8k-spk6135-pytorch ]; then |
|---|
| | | git lfs install |
|---|
