API tools faq
paste
Advertisement
Guest User

Untitled

a guest
Mar 21st, 2019
49
0
Never
Add comment
Not a member of Pastebin yet? Sign Up, it unlocks many cool features!
text 14.41 KB | None | 0 0
raw download clone embed print report
  1. SHELL := /bin/bash
  2.  
  3. # Use this file to override various settings
  4. -include /vagrant/options.aspire.Makefile
  5.  
  6. # Set to 'yes' if you want to do speaker ID for trs files
  7. DO_SPEAKER_ID?=no
  8. SID_THRESHOLD?=13
  9.  
  10. # Some audio produces no results (can't be segmented), or transcribes better
  11. # when segmented differently
  12. # Changing SEGMENTS to one of these values gives more flexibilty;
  13. # (see http://www-lium.univ-lemans.fr/diarization/doku.php/quick_start)
  14. #
  15. # show.seg : default - final segmentation with NCLR/CE clustering
  16. # show.i.seg : initial segmentation (entire audio)
  17. # show.pms.seg : sPeech/Music/Silence segmentation (don't use)
  18. # show.s.seg : GLR based segmentation, make small segments
  19. # show.l.seg : linear clustering (merge only side by side segments)
  20. # show.h.seg : hierarchical clustering
  21. # show.d.seg : viterbi decoding
  22. # show.adj.h.seg : boundaries adjusted
  23. # show.flt1.seg : filter spk segmentation according to pms segmentation
  24. # show.flt2.seg : filter spk segmentation according to pms segmentation
  25. # show.spl.seg : segments longer than 20 sec are split
  26. # show.spl10.seg : segments longer than 10 sec are split
  27. # show.g.seg : the gender and the bandwith are detected
  28. SEGMENTS ?= show.seg
  29.  
  30. # Where is Kaldi root directory?
  31. EESEN_ROOT ?= ~/eesen
  32. KALDI_ROOT ?=../kaldi
  33.  
  34. GRAPH_DIR?=$(EESEN_ROOT)/asr_egs/tedlium/v1/data/lang_phn_test
  35. MODEL_DIR?=$(EESEN_ROOT)/asr_egs/tedlium/v1/exp/train_phn_l5_c320
  36.  
  37. # How many processes to use for one transcription task
  38. # must be less than number of speakers, which for lium segmentation is often only 1
  39. njobs ?= 1
  40.  
  41. # How many threads to use in each process
  42. #nthreads ?= 1
  43.  
  44. # add Kaldi binaries to path
  45. PATH := utils:$(KALDI_ROOT)/src/bin:$(KALDI_ROOT)/tools/openfst/bin:$(KALDI_ROOT)/src/fstbin/:$(KALDI_ROOT)/src/gmmbin/:$(KALDI_ROOT)/src/featbin/:$(KALDI_ROOT)/src/lm/:$(KALDI_ROOT)/src/sgmmbin/:$(KALDI_ROOT)/src/sgmm2bin/:$(KALDI_ROOT)/src/fgmmbin/:$(KALDI_ROOT)/src/latbin/:$(KALDI_ROOT)/src/nnetbin:$(KALDI_ROOT)/src/nnet2bin/:$(KALDI_ROOT)/src/kwsbin:$(KALDI_ROOT)/src/ivectorbin:$(PATH)
  46. # add EESEN binaries to path
  47. PATH := $(EESEN_ROOT)/src/decoderbin:$(EESEN_ROOT)/src/featbin:$(EESEN_ROOT)/src/nnetbin:$(PATH)
  48. export train_cmd=run.pl
  49. export decode_cmd=run.pl
  50. export cuda_cmd=run.pl
  51. export mkgraph_cmd=run.pl
  52.  
  53. # optimum experimentally determined LM weight for TEDLIUM data set
  54. # (produces lowest WER)
  55. LM_SCALE?=8
  56.  
  57. # Find out where this Makefile is located (this is not really needed)
  58. where-am-i = $(lastword $(MAKEFILE_LIST))
  59. THIS_DIR := $(shell dirname $(call where-am-i))
  60.  
  61. # This ends up just being a folder name for output
  62. FINAL_PASS=eesen
  63.  
  64. #LD_LIBRARY_PATH=$(KALDI_ROOT)/tools/openfst/lib
  65.  
  66. .SECONDARY:
  67. .DELETE_ON_ERROR:
  68.  
  69. export
  70.  
  71. # Call this (once) before using the system
  72. .init: .kaldi #.lang .composed_lms
  73.  
  74. .kaldi:
  75. rm -f steps utils
  76. ln -fs $(KALDI_ROOT)/egs/wsj/s5/steps
  77. ln -fs $(KALDI_ROOT)/egs/wsj/s5/utils
  78. ln -fs $(KALDI_ROOT)/egs/sre08/v1/sid
  79. mkdir -p src-audio
  80.  
  81. build/audio/base/%.wav: src-audio/%.sph
  82. mkdir -p `dirname $@`
  83. #sox $^ build/audio/base/$*.wav rate -v $(sample_rate) #channels 1
  84. /data/ASR5/fmetze/kaldi-latest/tools/sph2pipe_v2.5/sph2pipe -f wav $^ | sox -t wav - -c 1 build/audio/base/$*.wav rate -v $(sample_rate)
  85.  
  86. build/audio/base/%.wav: src-audio/%.wav
  87. mkdir -p `dirname $@`
  88. sox $^ -c 1 -b 16 build/audio/base/$*.wav rate -v $(sample_rate)
  89.  
  90. build/audio/base/%.wav: src-audio/%.WAV
  91. mkdir -p `dirname $@`
  92. sox $^ -c 1 -b 16 build/audio/base/$*.wav rate -v $(sample_rate)
  93.  
  94. build/audio/base/%.wav: src-audio/%.mp3
  95. mkdir -p `dirname $@`
  96. # sox $^ -c 1 build/audio/base/$*.wav rate -v $(sample_rate)
  97. # lame --resample $(sample_rate) --decode $^ build/audio/base/$*.8k.wav
  98. # sox build/audio/base/$*.8k.wav -c 1 build/audio/base/$*.wav rate -v $(sample_rate)
  99. ffmpeg -loglevel warning -i $^ -f sox - | sox -t sox - -c 1 -b 16 -r $(sample_rate) $@
  100.  
  101. build/audio/base/%.wav: src-audio/%.MP3
  102. mkdir -p `dirname $@`
  103. sox $^ -c 1 build/audio/base/$*.wav rate -v $(sample_rate)
  104.  
  105. build/audio/base/%.wav: src-audio/%.ogg
  106. mkdir -p `dirname $@`
  107. sox $^ -c 1 build/audio/base/$*.wav rate -v $(sample_rate)
  108.  
  109. build/audio/base/%.wav: src-audio/%.mp2
  110. mkdir -p `dirname $@`
  111. sox $^ -c 1 build/audio/base/$*.wav rate -v $(sample_rate)
  112.  
  113. build/audio/base/%.wav: src-audio/%.wma
  114. mkdir -p `dirname $@`
  115. ffmpeg -loglevel warning -i $^ -f sox - | sox -t sox - -c 1 -b 16 -r $(sample_rate) $@
  116.  
  117. build/audio/base/%.wav: src-audio/%.WMA
  118. mkdir -p `dirname $@`
  119. ffmpeg -loglevel warning -i $^ -f sox - | sox -t sox - -c 1 -b 16 -r $(sample_rate) $@
  120.  
  121. build/audio/base/%.wav: src-audio/%.m4a
  122. mkdir -p `dirname $@`
  123. avconv -i $^ -ac 1 -ar $(sample_rate) -y $@
  124. # ffmpeg -i $^ -f sox - | sox -t sox - -c 1 -b 16 $@ rate -v $(sample_rate)
  125.  
  126. build/audio/base/%.wav: src-audio/%.mp4
  127. mkdir -p `dirname $@`
  128. # sox $^ -c 1 build/audio/base/$*.wav rate -v $(sample_rate)
  129. # avconv -i $^ -ac 1 -ar $(sample_rate) -y $@
  130. ffmpeg -i $^ -f sox - | sox -t sox - -c 1 -b 16 $@ rate -v $(sample_rate)
  131. echo "converted audio"
  132. date +%s%N | cut -b1-13
  133.  
  134. build/audio/base/%.wav: src-audio/%.MP4
  135. mkdir -p `dirname $@`
  136. # sox $^ -c 1 build/audio/base/$*.wav rate -v $(sample_rate)
  137. avconv -i $^ -ac 1 -ar $(sample_rate) -y $@
  138. echo "converted audio"
  139. date +%s%N | cut -b1-13
  140.  
  141. build/audio/base/%.wav: src-audio/%.flac
  142. mkdir -p `dirname $@`
  143. flac -s -d -c $^ | sox -t wav - -c 1 -b 16 build/audio/base/$*.wav rate -v $(sample_rate)
  144.  
  145. build/audio/base/%.wav: src-audio/%.amr
  146. mkdir -p `dirname $@`
  147. amrnb-decoder $^ $@.tmp.raw
  148. sox -s -b 16 -c 1 -r 8000 $@.tmp.raw -c 1 build/audio/base/$*.wav rate -v $(sample_rate)
  149. rm $@.tmp.raw
  150.  
  151. build/audio/base/%.wav: src-audio/%.mpg
  152. mkdir -p `dirname $@`
  153. avconv -i $^ -f sox - | sox -t sox - -c 1 -b 16 build/audio/base/$*.wav rate -v $(sample_rate)
  154. # ffmpeg -i $^ -f sox - | sox -t sox - -c 1 -b 16 build/audio/base/$*.wav rate -v $(sample_rate)
  155.  
  156. # Speaker diarization
  157. build/diarization/%/$(SEGMENTS): build/audio/base/%.wav
  158. rm -rf `dirname $@`
  159. mkdir -p `dirname $@`
  160. echo "$* 1 0 1000000000 U U U 1" > `dirname $@`/show.uem.seg;
  161. ./scripts/diarization.sh $^ `dirname $@`/show.uem.seg $(SEGMENTS);
  162. echo "diarization complete"
  163. date +%s%N | cut -b1-13
  164.  
  165. #build/audio/segmented/%: build/diarization/%/show.seg
  166. build/audio/segmented/%: build/diarization/%/$(SEGMENTS)
  167. rm -rf $@
  168. mkdir -p $@
  169. cat $^ | cut -f 3,4,8 -d " " | \
  170. while read LINE ; do \
  171. start=`echo $$LINE | cut -f 1 -d " " | perl -npe '$$_=$$_/100.0'`; \
  172. len=`echo $$LINE | cut -f 2 -d " " | perl -npe '$$_=$$_/100.0'`; \
  173. sp_id=`echo $$LINE | cut -f 3 -d " "`; \
  174. timeformatted=`echo "$$start $$len" | perl -ne '@t=split(); $$start=$$t[0]; $$len=$$t[1]; $$end=$$start+$$len; printf("%08.3f-%08.3f\n", $$start,$$end);'` ; \
  175. if [ $${sp_id} == 'A' ]; then \
  176. sox build/audio/base/$*.wav -c 1 $@/$*_$${timeformatted}_$${sp_id}.wav trim $$start $$len remix 1; \
  177. elif [ $${sp_id} == 'B' ]; then \
  178. sox build/audio/base/$*.wav -c 1 $@/$*_$${timeformatted}_$${sp_id}.wav trim $$start $$len remix 2; \
  179. else \
  180. sox build/audio/base/$*.wav --norm $@/$*_$${timeformatted}_$${sp_id}.wav trim $$start $$len; \
  181. fi \
  182. done
  183.  
  184. build/trans/%/wav.scp: build/audio/segmented/%
  185. mkdir -p `dirname $@`
  186. /bin/ls $</*.wav | \
  187. perl -npe 'chomp; $$orig=$$_; s/.*\/(.*)_(\d+\.\d+-\d+\.\d+)_(.*)\.wav/\1-\3---\2/; $$_=$$_ . " $$orig\n";' | LC_ALL=C sort > $@
  188.  
  189. build/trans/%/utt2spk: build/trans/%/wav.scp
  190. cat $^ | perl -npe 's/\s+.*//; s/((.*)---.*)/\1 \2/' > $@
  191.  
  192. build/trans/%/spk2utt: build/trans/%/utt2spk
  193. utils/utt2spk_to_spk2utt.pl $^ > $@
  194.  
  195.  
  196. # FBANK calculation
  197. # example target:
  198. # make build/trans/myvideo/fbank
  199. # note the % pattern matches e.g. myvideo
  200. build/trans/%/fbank: build/trans/%/spk2utt
  201. rm -rf $@ build/trans/$*/cmvn.scp
  202. steps/$(fbank).sh --fbank-config conf/fbank.$(sample_rate).conf --cmd "$$train_cmd" --nj $(njobs) \
  203. build/trans/$* build/trans/$*/exp/make_fbank $@ || exit 1
  204. steps/compute_cmvn_stats.sh build/trans/$* build/trans/$*/exp/make_fbank $@ || exit 1;
  205.  
  206. # Hires MFCC calculation
  207. build/trans/%/mfcc_hires: build/trans/%/spk2utt
  208. rm -rf $@ build/trans/$*/cmvn.scp
  209. steps/make_mfcc.sh --mfcc-config conf/mfcc_hires.conf --cmd "$$train_cmd" --nj $(njobs) \
  210. build/trans/$* build/trans/$*/exp/make_mfcc_hires $@ || exit 1
  211. steps/compute_cmvn_stats.sh build/trans/$* build/trans/$*/exp/make_mfcc_hires $@ || exit 1;
  212.  
  213. # i-vectors for adapted Aspire model and training
  214. # this should not be hardcoded in this way, but well ...
  215. aspireivectordir=/data/ASR5/sdalmia_1/spring2018/kaldi/kaldi/egs/aspire/s5_orig
  216. build/trans/%/ivectors: build/trans/%/mfcc_hires
  217. steps/online/nnet2/extract_ivectors.sh --cmd "$$train_cmd" --nj $(njobs) --silence-weight 0.00001 \
  218. --sub-speaker-frames 6000 --max-count 75 \
  219. build/trans/$* "$$aspireivectordir"/data/lang \
  220. "$$aspireivectordir"/exp/tdnn_7b_chain_online/ivector_extractor build/trans/$*/exp/make_ivectors || exit 1;
  221.  
  222.  
  223. # Decode
  224. # example target
  225. # make build/trans/myvideo/eesen/decode/log
  226. build/trans/%/$(FINAL_PASS)/decode/log: build/trans/%/spk2utt build/trans/%/fbank
  227. rm -rf build/trans/$*/$(FINAL_PASS) && mkdir -p build/trans/$*/$(FINAL_PASS)
  228. # (cd build/trans/$*/eesen; for f in $(MODEL_DIR)/*; do ln -s $$f; done)
  229. ifeq ($(MODELS),aspire.kaldi)
  230. local/decode_nnet3.sh --cmd "$$decode_cmd" --nj $(njobs) \
  231. --acwt $(ACWT) --post_decode_acwt 10.0 --beam $(BEAM) \
  232. --lattice_beam 6.0 --max-active 7000 --skip_scoring true \
  233. $(GRAPH_DIR) build/trans/$* `dirname $@` $(MODEL_DIR) || exit 1;
  234. else ifeq ($(MODELS),tedlium.eesen)
  235. $(EESEN_ROOT)/local/decode_ctc_lat_tf.sh --mdl $(MODEL_DIR) --scoredir build/trans/$* \
  236. --cmd "$$decode_cmd" --nj $(njobs) --beam $(BEAM) \
  237. --lattice_beam 8.0 --max-active 5000 --skip_scoring true \
  238. --acwt $(ACWT) --temperature 1.25 \
  239. $(GRAPH_DIR) build/trans/$* `dirname $@` || exit 1;
  240. endif
  241.  
  242. # scoring can happen here now, get_ctm_conf.sh only scores if -f build/trans/$*/stm
  243. # produces confidence scores
  244. # e.g. make build/trans/myvideo/eesen/decode/.ctm
  245. # % = build/trans/myvideo/eesen
  246. %/decode/.ctm: %/decode/log
  247. # local/get_ctm.sh `dirname $*` $*/graph $*/decode
  248. # local/get_ctm_conf.sh `dirname $*` $*/graph $*/decode
  249. ifeq ($(MODELS),aspire.kaldi)
  250. local/lattice_to_ctm_aspire.sh `dirname $*` $(GRAPH_DIR) $*/decode
  251. else
  252. local/get_ctm_conf.sh `dirname $*` $(GRAPH_DIR) $*/decode
  253. endif
  254. touch -m $@
  255.  
  256. # % = myvideo/eesen
  257. # e.g. make build/trans/myvideo/eesen.segmented.splitw2.ctm
  258. build/trans/%.segmented.splitw2.ctm: build/trans/%/decode/.ctm
  259. # cat build/trans/$*/decode/score_$(LM_SCALE)/`dirname $*`.ctm | perl -npe 's/(.*)-(S\d+)---(\S+)/\1_\3_\2/' > $@
  260. cat build/trans/$*/decode/score_$(LM_SCALE)/`dirname $*`.ctm | perl -npe 's/(.*)-(\w+)---(\S+)/\1_\3_\2/' > $@
  261.  
  262.  
  263. #build/trans/myvideo/eesen.segmented.splitw2.ctm -> build/trans/myvideo/eesen.segmented.with-compounds.ctm
  264. %.with-compounds.ctm: %.splitw2.ctm
  265. cat $*.splitw2.ctm > $@
  266.  
  267. #build/trans/myvideo/eesen.segmented.with-compounds.ctm -> build/trans/myvideo/eesen.segmented.ctm
  268. %.segmented.ctm: %.segmented.with-compounds.ctm
  269. cat $^ > $@
  270.  
  271. %.ctm: %.segmented.ctm
  272. cat $^ | python scripts/unsegment-ctm.py | LC_ALL=C sort -k 1,1 -k 3,3n -k 4,4n > $@
  273.  
  274. %.with-compounds.ctm: %.segmented.with-compounds.ctm
  275. cat $^ | python scripts/unsegment-ctm.py | LC_ALL=C sort -k 1,1 -k 3,3n -k 4,4n > $@
  276.  
  277. %.hyp: %.segmented.ctm
  278. cat $^ | python scripts/segmented-ctm-to-hyp.py > $@
  279.  
  280. ifeq "yes" "$(DO_SPEAKER_ID)"
  281. build/trans/%/$(FINAL_PASS).trs: build/trans/%/$(FINAL_PASS).hyp build/trans/%/sid-result.txt
  282. cat build/trans/$*/$(FINAL_PASS).hyp | python scripts/hyp2trs.py --sid build/trans/$*/sid-result.txt > $@
  283. else
  284. build/trans/%/$(FINAL_PASS).trs: build/trans/%/$(FINAL_PASS).hyp
  285. cat $^ | python scripts/hyp2trs.py > $@
  286. endif
  287.  
  288. %.sbv: %.hyp
  289. cat $^ | python scripts/hyp2sbv.py > $@
  290.  
  291. %.txt: %.hyp
  292. cat $^ | perl -npe 'use locale; s/ \(\S+\)/\./; $$_= ucfirst();' > $@
  293.  
  294. %.srt: %.ctm
  295. cat $^ | python scripts/ctm2srt.py > $@
  296.  
  297. %.labels: %.ctm
  298. cat $^ | python scripts/ctm2labels.py > $@
  299.  
  300. build/output/%.trs: build/trans/%/$(FINAL_PASS).trs
  301. mkdir -p `dirname $@`
  302. cp $^ $@
  303.  
  304. build/output/%.ctm: build/trans/%/$(FINAL_PASS).ctm
  305. mkdir -p `dirname $@`
  306. cp $^ $@
  307.  
  308. build/output/%.txt: build/trans/%/$(FINAL_PASS).txt
  309. mkdir -p `dirname $@`
  310. cp $^ $@
  311.  
  312. build/output/%.with-compounds.ctm: build/trans/%/$(FINAL_PASS).with-compounds.ctm
  313. mkdir -p `dirname $@`
  314. cp $^ $@
  315.  
  316. build/output/%.sbv: build/trans/%/$(FINAL_PASS).sbv
  317. mkdir -p `dirname $@`
  318. cp $^ $@
  319.  
  320. build/output/%.srt: build/trans/%/$(FINAL_PASS).ctm
  321. mkdir -p `dirname $@`
  322. cp $^ $@
  323.  
  324. build/output/%.ali: build/trans/%/$(FINAL_PASS).txt
  325. mkdir -p `dirname $@`
  326. ./run_align.sh --GRAPH_DIR $(GRAPH_DIR) --MODEL_DIR $(MODEL_DIR) $*
  327.  
  328. ### Speaker ID stuff
  329. # i-vectors for each speaker in our audio file
  330. build/trans/%/ivectors: build/trans/%/mfcc
  331. sid/extract_ivectors.sh --cmd "$$decode_cmd" --nj $(njobs) \
  332. $(THIS_DIR)/kaldi-data/extractor_2048_top500 build/trans/$* $@
  333.  
  334. # a cross product of train and test speakers
  335. build/trans/%/sid-trials.txt: build/trans/%/ivectors
  336. cut -f 1 -d " " $(THIS_DIR)/kaldi-data/ivectors_train_top500/spk_ivector.scp | \
  337. while read a; do \
  338. cut -f 1 -d " " build/trans/$*/ivectors/spk_ivector.scp | \
  339. while read b; do \
  340. echo "$$a $$b"; \
  341. done ; \
  342. done > $@
  343.  
  344. # similarity scores
  345. build/trans/%/sid-scores.txt: build/trans/%/sid-trials.txt
  346. ivector-plda-scoring \
  347. "ivector-copy-plda --smoothing=0.0 $(THIS_DIR)/kaldi-data/ivectors_train_top500/plda - |" \
  348. "ark:ivector-subtract-global-mean scp:$(THIS_DIR)/kaldi-data/ivectors_train_top500/spk_ivector.scp ark:- |" \
  349. "ark:ivector-subtract-global-mean scp:build/trans/$*/ivectors/spk_ivector.scp ark:- |" \
  350. build/trans/$*/sid-trials.txt $@
  351.  
  352. # pick speakers above the threshold
  353. build/trans/%/sid-result.txt: build/trans/%/sid-scores.txt
  354. cat build/trans/$*/sid-scores.txt | sort -u -k 2,2 -k 3,3nr | sort -u -k2,2 | \
  355. awk 'int($$3)>=$(SID_THRESHOLD)' | perl -npe 's/(\S+) \S+-(S\d+) \S+/\2 \1/; s/-/ /g' > $@
  356.  
  357.  
  358. # Meta-target that deletes all files created during processing a file. Call e.g. 'make .etteytlus2013.clean
  359. .%.clean:
  360. rm -rf build/audio/base/$*.wav build/audio/segmented/$* build/diarization/$* build/trans/$*
  361. # rm -rf build/audio/base/$*.wav build/audio/segmented/$* build/diarization/$* build/trans/$* #src-audio/$*.wav
  362.  
  363. # Also deletes the output files
  364. .%.cleanest: .%.clean
  365. rm -rf build/output/$*.{trs,txt,ctm,with-compounds.ctm,sbv,ali,labels}
Advertisement
Add Comment
Please, Sign In to add comment
Advertisement
Public Pastes
Advertisement
We use cookies for various purposes including analytics. By continuing to use Pastebin, you agree to our use of cookies as described in the Cookies Policy.  OK, I Understand
Not a member of Pastebin yet?
Sign Up, it unlocks many cool features!