def test_sick_lda_mllt(sick_dict, sick_corpus, generated_dir): a = TrainableAligner(sick_corpus, sick_dict, os.path.join(generated_dir, 'sick_output'), temp_directory=os.path.join(generated_dir, 'sickcorpus')) a.train_lda_mllt() a.export_textgrids()
def align_corpus(args): if not args.temp_directory: temp_dir = TEMP_DIR else: temp_dir = os.path.expanduser(args.temp_directory) corpus_name = os.path.basename(args.corpus_directory) if corpus_name == '': args.corpus_directory = os.path.dirname(args.corpus_directory) corpus_name = os.path.basename(args.corpus_directory) data_directory = os.path.join(temp_dir, corpus_name) conf_path = os.path.join(data_directory, 'config.yml') if os.path.exists(conf_path): with open(conf_path, 'r') as f: conf = yaml.load(f) else: conf = { 'dirty': False, 'begin': time.time(), 'version': __version__, 'type': 'train_and_align', 'corpus_directory': args.corpus_directory, 'dictionary_path': args.dictionary_path } if getattr(args, 'clean', False) \ or conf['dirty'] or conf['type'] != 'train_and_align' \ or conf['corpus_directory'] != args.corpus_directory \ or conf['version'] != __version__ \ or conf['dictionary_path'] != args.dictionary_path: shutil.rmtree(data_directory, ignore_errors=True) shutil.rmtree(args.output_directory, ignore_errors=True) os.makedirs(data_directory, exist_ok=True) os.makedirs(args.output_directory, exist_ok=True) try: corpus = Corpus(args.corpus_directory, data_directory, speaker_characters=args.speaker_characters, num_jobs=getattr(args, 'num_jobs', 3), debug=getattr(args, 'debug', False), ignore_exceptions=getattr(args, 'ignore_exceptions', False)) dictionary = Dictionary(args.dictionary_path, data_directory, word_set=corpus.word_set) utt_oov_path = os.path.join(corpus.split_directory, 'utterance_oovs.txt') if os.path.exists(utt_oov_path): shutil.copy(utt_oov_path, args.output_directory) oov_path = os.path.join(corpus.split_directory, 'oovs_found.txt') if os.path.exists(oov_path): shutil.copy(oov_path, args.output_directory) mono_params = {'align_often': not args.fast} tri_params = {'align_often': not args.fast} tri_fmllr_params = {'align_often': not args.fast} a = TrainableAligner(corpus, dictionary, args.output_directory, temp_directory=data_directory, mono_params=mono_params, tri_params=tri_params, tri_fmllr_params=tri_fmllr_params, num_jobs=args.num_jobs, skip_input=getattr(args, 'quiet', False), nnet=getattr(args, 'artificial_neural_net', False)) a.verbose = args.verbose # GMM training (looks like it needs to be done either way, as a starter for nnet) a.train_mono() a.export_textgrids() a.train_tri() a.export_textgrids() a.train_tri_fmllr() a.export_textgrids() # nnet training if args.artificial_neural_net: # Do nnet training a.train_lda_mllt() #a.train_diag_ubm() # Uncomment to train i-vector extractor #a.ivector_extractor() # Uncomment to train i-vector extractor (integrate with argument eventually) a.train_nnet_basic() a.export_textgrids() if args.output_model_path is not None: a.save(args.output_model_path) except: conf['dirty'] = True raise finally: with open(conf_path, 'w') as f: yaml.dump(conf, f)