Optimized speaker-based processing in subsets

montreal_forced_aligner/command_line/align.py

@@ -12,7 +12,7 @@
 from montreal_forced_aligner.config import TEMP_DIR, align_yaml_to_config, load_basic_align
 from montreal_forced_aligner.utils import get_available_acoustic_languages, get_pretrained_acoustic_path, \
     get_available_dict_languages, get_dictionary_path
-from montreal_forced_aligner.helper import setup_logger
+from montreal_forced_aligner.helper import setup_logger, log_config
 from montreal_forced_aligner.exceptions import ArgumentError
 
 
@@ -39,6 +39,8 @@ def align_corpus(args, unknown_args=None):
         print('Cleaning old directory!')
         shutil.rmtree(data_directory, ignore_errors=True)
     logger = setup_logger(command, data_directory)
+    logger.debug('ALIGN CONFIG:')
+    log_config(logger, align_config)
     if os.path.exists(conf_path):
         with open(conf_path, 'r') as f:
             conf = yaml.load(f, Loader=yaml.SafeLoader)

montreal_forced_aligner/command_line/train_and_align.py

@@ -10,7 +10,7 @@
 from montreal_forced_aligner.aligner import TrainableAligner
 from montreal_forced_aligner.config import TEMP_DIR, train_yaml_to_config, load_basic_train
 from montreal_forced_aligner.utils import get_available_dict_languages, get_dictionary_path
-from montreal_forced_aligner.helper import setup_logger
+from montreal_forced_aligner.helper import setup_logger, log_config
 
 from montreal_forced_aligner.exceptions import ArgumentError
 
@@ -38,6 +38,10 @@ def align_corpus(args, unknown_args=None):
         print('Cleaning old directory!')
         shutil.rmtree(data_directory, ignore_errors=True)
     logger = setup_logger(command, data_directory)
+    logger.debug('TRAIN CONFIG:')
+    log_config(logger, train_config)
+    logger.debug('ALIGN CONFIG:')
+    log_config(logger, align_config)
     if args.debug:
         logger.warning('Running in DEBUG mode, may have impact on performance and disk usage.')
     if os.path.exists(conf_path):

montreal_forced_aligner/command_line/train_dictionary.py

@@ -12,7 +12,7 @@
 from montreal_forced_aligner.config import TEMP_DIR, align_yaml_to_config, load_basic_align
 from montreal_forced_aligner.utils import get_available_acoustic_languages, get_pretrained_acoustic_path, \
     get_available_dict_languages, get_dictionary_path
-from montreal_forced_aligner.helper import setup_logger
+from montreal_forced_aligner.helper import setup_logger, log_config
 from montreal_forced_aligner.exceptions import ArgumentError
 
 
@@ -37,6 +37,8 @@ def train_dictionary(args):
         print('Cleaning old directory!')
         shutil.rmtree(data_directory, ignore_errors=True)
     logger = setup_logger(command, data_directory)
+    logger.debug('ALIGN CONFIG:')
+    log_config(logger, align_config)
     if os.path.exists(conf_path):
         with open(conf_path, 'r') as f:
             conf = yaml.load(f, Loader=yaml.SafeLoader)

montreal_forced_aligner/command_line/train_ivector_extractor.py

@@ -9,7 +9,7 @@
 from montreal_forced_aligner.dictionary import Dictionary
 from montreal_forced_aligner.aligner import PretrainedAligner
 from montreal_forced_aligner.config import TEMP_DIR, train_yaml_to_config, load_basic_train_ivector
-from montreal_forced_aligner.helper import setup_logger
+from montreal_forced_aligner.helper import setup_logger, log_config
 from montreal_forced_aligner.utils import get_available_acoustic_languages, get_pretrained_acoustic_path, \
     get_available_dict_languages, get_dictionary_path
 from montreal_forced_aligner.models import AcousticModel
@@ -38,6 +38,10 @@ def train_ivector(args):
         print('Cleaning old directory!')
         shutil.rmtree(data_directory, ignore_errors=True)
     logger = setup_logger(command, data_directory)
+    logger.debug('TRAIN CONFIG:')
+    log_config(logger, train_config)
+    logger.debug('ALIGN CONFIG:')
+    log_config(logger, align_config)
 
     if os.path.exists(conf_path):
         with open(conf_path, 'r') as f:

montreal_forced_aligner/command_line/transcribe.py

@@ -9,7 +9,7 @@
 from montreal_forced_aligner.dictionary import Dictionary
 from montreal_forced_aligner.transcriber import Transcriber
 from montreal_forced_aligner.models import AcousticModel, LanguageModel, FORMAT
-from montreal_forced_aligner.helper import setup_logger
+from montreal_forced_aligner.helper import setup_logger, log_config
 from montreal_forced_aligner.config import TEMP_DIR, transcribe_yaml_to_config, load_basic_transcribe, save_config
 from montreal_forced_aligner.utils import get_available_acoustic_languages, get_pretrained_acoustic_path, \
     get_available_lm_languages, get_pretrained_language_model_path, \
@@ -37,6 +37,8 @@ def transcribe_corpus(args):
         print('Cleaning old directory!')
         shutil.rmtree(data_directory, ignore_errors=True)
     logger = setup_logger(command, data_directory)
+    logger.debug('TRANSCRIBE CONFIG:')
+    log_config(logger, transcribe_config)
     os.makedirs(data_directory, exist_ok=True)
     os.makedirs(args.output_directory, exist_ok=True)
     os.makedirs(data_directory, exist_ok=True)

montreal_forced_aligner/config/align_config.py

@@ -41,6 +41,7 @@ def update(self, data):
 
     def update_from_args(self, args):
         super(AlignConfig, self).update_from_args(args)
+        self.feature_config.update_from_args(args)
         if self.retry_beam <= self.beam:
             self.retry_beam = self.beam * 4
 

montreal_forced_aligner/corpus/align_corpus.py

@@ -374,12 +374,6 @@ def check_warnings(self):
 
         if len(self.speak_utt_mapping) < self.num_jobs:
             self.num_jobs = len(self.speak_utt_mapping)
-        if self.num_jobs < len(self.sample_rates.keys()):
-            self.num_jobs = len(self.sample_rates.keys())
-            msg = 'The number of jobs was set to {}, due to the different sample rates in the dataset. ' \
-                  'If you would like to use fewer parallel jobs, ' \
-                  'please resample all wav files to the same sample rate.'.format(self.num_jobs)
-            self.logger.warning(msg)
 
     def save_text_file(self, file_name):
         if self.segments:
@@ -617,6 +611,7 @@ def create_subset(self, subset, feature_config):
         split_directory = self.split_directory()
         subset_directory = os.path.join(self.output_directory, 'subset_{}'.format(subset))
         subset_utt_path = os.path.join(subset_directory, 'included_utts.txt')
+        subset_speaker_path = os.path.join(subset_directory, 'included_speakers.txt')
         if os.path.exists(subset_utt_path):
             subset_utts = []
             with open(subset_utt_path, 'r', encoding='utf8') as f:
@@ -638,7 +633,14 @@ def create_subset(self, subset, feature_config):
             with open(subset_utt_path, 'w', encoding='utf8') as f:
                 for u in subset_utts:
                     f.write('{}\n'.format(u))
-        for i, s in enumerate(self.speakers):
+        subset_speakers = set()
+        for u in subset_utts:
+            subset_speakers.add(self.utt_speak_mapping[u])
+        subset_speakers = sorted(subset_speakers)
+        with open(subset_speaker_path, 'w') as f:
+            for s in subset_speakers:
+                f.write(s + '\n')
+        for i, s in enumerate(subset_speakers):
             for fn in ['text.{}', 'text.{}.int', 'utt2spk.{}']:
                 sub_path = os.path.join(subset_directory, fn.format(i))
                 with open(os.path.join(split_directory, fn.format(i)), 'r', encoding='utf8') as inf, \
@@ -648,7 +650,6 @@ def create_subset(self, subset, feature_config):
                         if s[0] not in subset_utts:
                             continue
                         outf.write(line)
-            subset_speakers = []
             sub_path = os.path.join(subset_directory, 'spk2utt.{}'.format(i))
             with open(os.path.join(split_directory, 'spk2utt.{}'.format(i)), 'r', encoding='utf8') as inf, \
                     open(sub_path, 'w', encoding='utf8') as outf:
@@ -659,7 +660,6 @@ def create_subset(self, subset, feature_config):
                     if not filtered_utts:
                         continue
                     outf.write('{} {}\n'.format(speaker, ' '.join(filtered_utts)))
-                    subset_speakers.append(speaker)
             sub_path = os.path.join(subset_directory, 'cmvn.{}.scp'.format(i))
             with open(os.path.join(split_directory, 'cmvn.{}.scp'.format(i)), 'r', encoding='utf8') as inf, \
                     open(sub_path, 'w', encoding='utf8') as outf:

montreal_forced_aligner/corpus/base.py

@@ -199,6 +199,8 @@ def parse_features_logs(self):
 
     def speaker_utterance_info(self):
         num_speakers = len(self.speak_utt_mapping.keys())
+        if not num_speakers:
+            raise CorpusError('Could not find any utterances in {}'.format(self.directory))
         average_utterances = sum(len(x) for x in self.speak_utt_mapping.values()) / num_speakers
         msg = 'Number of speakers in corpus: {}, average number of utterances per speaker: {}'.format(num_speakers,
                                                                                                       average_utterances)

montreal_forced_aligner/features/config.py

@@ -5,6 +5,7 @@
 from .processing import mfcc, compute_vad
 
 from ..helper import thirdparty_binary, load_scp, save_speaker_groups
+from ..config.base_config import BaseConfig
 
 
 def make_safe(value):
@@ -13,7 +14,7 @@ def make_safe(value):
     return str(value)
 
 
-class FeatureConfig(object):
+class FeatureConfig(BaseConfig):
     """
     Class to store configuration information about MFCC generation
 

montreal_forced_aligner/helper.py

@@ -4,7 +4,7 @@
 from typing import Any, List, Tuple
 import logging
 import sys
-
+import yaml
 from .exceptions import ThirdpartyError, KaldiProcessingError
 
 Labels = List[Any]
@@ -198,6 +198,10 @@ def setup_logger(identifier, output_directory):
 
     return logger
 
+def log_config(logger, config):
+
+    stream = yaml.dump(config)
+    logger.debug(stream)
 
 def parse_logs(log_directory):
     error_logs = []

montreal_forced_aligner/multiprocessing/alignment.py

@@ -551,6 +551,7 @@ def ali_to_textgrid_func(model_directory, word_path, split_directory, job_name,
 
         subprocess.call([thirdparty_binary('nbest-to-ctm'),
                          '--frame-shift={}'.format(frame_shift),
+                         '--precision=3',
                          'ark:' + aligned_path,
                          word_ctm_path],
                         stderr=log_file)
@@ -560,6 +561,7 @@ def ali_to_textgrid_func(model_directory, word_path, split_directory, job_name,
                                       stderr=log_file)
         nbest_proc = subprocess.Popen([thirdparty_binary('nbest-to-ctm'),
                                        '--frame-shift={}'.format(frame_shift),
+                                       '--precision=3',
                                        "ark:-", phone_ctm_path],
                                       stdin=phone_proc.stdout,
                                       stderr=log_file)
@@ -633,7 +635,7 @@ def convert_ali_to_textgrids(align_config, output_directory, model_directory, di
                 phone_ctm[k] = v
             else:
                 phone_ctm[k].update(v)
-    ctm_to_textgrid(word_ctm, phone_ctm, output_directory, corpus, dictionary)
+    ctm_to_textgrid(word_ctm, phone_ctm, output_directory, corpus, dictionary, frameshift=frame_shift)
 
 
 def tree_stats_func(directory, ci_phones, mdl, feature_string, ali_path, job_name):
@@ -685,10 +687,15 @@ def tree_stats(directory, align_directory, split_directory, ci_phones, speakers,
     tree_accs = [os.path.join(directory, '{}.treeacc'.format(x)) for x in range(len(speakers))]
     log_path = os.path.join(directory, 'log', 'sum_tree_acc.log')
     with open(log_path, 'w', encoding='utf8') as log_file:
-        subprocess.call([thirdparty_binary('sum-tree-stats'), os.path.join(directory, 'treeacc')] +
-                        tree_accs, stderr=log_file)
-    for f in tree_accs:
-        os.remove(f)
+        tmp_stats_path = os.path.join(directory, 'tempacc')
+        output_stats_path = os.path.join(directory, 'treeacc')
+        for af in tree_accs:
+            if not os.path.exists(output_stats_path):
+                os.rename(af, output_stats_path)
+                continue
+            subprocess.call([thirdparty_binary('sum-tree-stats'), tmp_stats_path, output_stats_path, af], stderr=log_file)
+            os.remove(af)
+            os.rename(tmp_stats_path, output_stats_path)
 
 
 def convert_alignments_func(directory, align_directory, job_name):

montreal_forced_aligner/textgrid.py

@@ -70,14 +70,23 @@ def ctm_to_textgrid(word_ctm, phone_ctm, out_directory, corpus, dictionary, fram
                 wordtier = IntervalTier(name='words', maxTime=maxtime)
                 phonetier = IntervalTier(name='phones', maxTime=maxtime)
                 phonetier_len = len(phone_ctm[k][speaker])
-                for interval in v:
+                for interval in sorted(v):
                     if maxtime - interval[1] < frameshift:  # Fix rounding issues
                         interval[1] = maxtime
-                    wordtier.add(*interval)
+                    try:
+                        wordtier.add(*interval)
+                    except ValueError:
+
+                        interval[0] = wordtier[-1].maxTime
+                        wordtier.add(*interval)
                 for j, interval in enumerate(phone_ctm[k][speaker]):
                     if j == phonetier_len - 1:  # sync last phone boundary to end of audio file
                         interval[1] = maxtime
-                    phonetier.add(*interval)
+                    try:
+                        phonetier.add(*interval)
+                    except ValueError:
+                        interval[0] = phonetier[-1].maxTime
+                        phonetier.add(*interval)
                 tg.append(wordtier)
                 tg.append(phonetier)
                 relative = corpus.file_directory_mapping[k]

montreal_forced_aligner/trainers/base.py

@@ -209,6 +209,15 @@ def align(self, subset, call_back=None):
                 align_data_directory = self.corpus.split_directory()
             else:
                 align_data_directory = self.corpus.subset_directory(subset, self.feature_config)
+
+            subset_speaker_path = os.path.join(align_data_directory, 'included_speakers.txt')
+            if os.path.exists(subset_speaker_path):
+                speakers = []
+                with open(subset_speaker_path, 'r', encoding='utf8') as f:
+                    for line in f:
+                        speakers.append(line.strip())
+            else:
+                speakers = self.corpus.speakers
             try:
                 log_dir = os.path.join(self.align_directory, 'log')
                 os.makedirs(log_dir, exist_ok=True)
@@ -222,10 +231,10 @@ def align(self, subset, call_back=None):
                 shutil.copyfile(os.path.join(self.train_directory, 'final.occs'),
                                 os.path.join(self.align_directory, 'final.occs'))
                 compile_train_graphs(self.align_directory, self.dictionary.output_directory,
-                                     align_data_directory, self.corpus.speakers, self.corpus.num_jobs, self)
+                                     align_data_directory, speakers, self.corpus.num_jobs, self)
                 align('final', self.align_directory, align_data_directory,
                       self.dictionary.optional_silence_csl,
-                      self.corpus.speakers, self.corpus.num_jobs, self, self.align_directory)
+                      speakers, self.corpus.num_jobs, self, self.align_directory)
                 self.save(os.path.join(self.align_directory, 'acoustic_model.zip'))
             except Exception as e:
                 with open(dirty_path, 'w'):
@@ -254,6 +263,15 @@ def train(self, call_back=None):
             iters = tqdm(range(1, self.num_iterations))
         else:
             iters = range(1, self.num_iterations)
+
+        subset_speaker_path = os.path.join(self.data_directory, 'included_speakers.txt')
+        if os.path.exists(subset_speaker_path):
+            speakers = []
+            with open(subset_speaker_path, 'r', encoding='utf8') as f:
+                for line in f:
+                    speakers.append(line.strip())
+        else:
+            speakers = self.corpus.speakers
         try:
             for i in iters:
                 model_path = os.path.join(self.train_directory, '{}.mdl'.format(i))
@@ -264,31 +282,40 @@ def train(self, call_back=None):
                 if i in self.realignment_iterations:
                     align(i, self.train_directory, self.data_directory,
                           self.dictionary.optional_silence_csl,
-                          self.corpus.speakers, self.corpus.num_jobs, self)
+                          speakers, self.corpus.num_jobs, self)
                     if self.debug:
-                        compute_alignment_improvement(i, self, self.train_directory, self.corpus.speakers,
+                        compute_alignment_improvement(i, self, self.train_directory, speakers,
                                                       self.corpus.num_jobs)
-                acc_stats(i, self.train_directory, self.data_directory, self.corpus.speakers, self.corpus.num_jobs,
+                acc_stats(i, self.train_directory, self.data_directory, speakers, self.corpus.num_jobs,
                           self)
                 log_path = os.path.join(self.log_directory, 'update.{}.log'.format(i))
-                with open(log_path, 'w') as logf:
+                with open(log_path, 'w') as log_file:
                     acc_files = [os.path.join(self.train_directory, '{}.{}.acc'.format(i, x))
-                                 for x in range(len(self.corpus.speakers))]
+                                 for x in range(len(speakers))]
+                    tmp_stats_path = os.path.join(self.train_directory, 'temp.acc')
+                    output_stats_path = os.path.join(self.train_directory, '{}.acc'.format(i))
+                    for af in acc_files:
+                        if not os.path.exists(output_stats_path):
+                            os.rename(af, output_stats_path)
+                            continue
+                        sum_proc = subprocess.Popen([thirdparty_binary('gmm-sum-accs'),
+                                                     tmp_stats_path,
+                                                     output_stats_path, af],
+                                                stderr=log_file)
+                        sum_proc.communicate()
+                        os.remove(af)
+                        os.rename(tmp_stats_path, output_stats_path)
                     est_proc = subprocess.Popen([thirdparty_binary('gmm-est'),
                                                  '--write-occs=' + occs_path,
                                                  '--mix-up=' + str(num_gauss), '--power=' + str(self.power),
                                                  model_path,
-                                                 "{} - {}|".format(thirdparty_binary('gmm-sum-accs'),
-                                                                   ' '.join(map(make_path_safe, acc_files))),
+                                                 output_stats_path,
                                                  next_model_path],
-                                                stderr=logf)
+                                                stderr=log_file)
                     est_proc.communicate()
-                if not self.debug:
-                    for f in acc_files:
-                        os.remove(f)
                 if not os.path.exists(next_model_path):
                     raise (Exception('There was an error training in iteration {}, please check the logs.'.format(i)))
-                self.parse_log_directory(self.log_directory, i, self.corpus.speakers, call_back)
+                self.parse_log_directory(self.log_directory, i, speakers, call_back)
                 if i < self.final_gaussian_iteration:
                     num_gauss += self.gaussian_increment
             shutil.copy(os.path.join(self.train_directory, '{}.mdl'.format(self.num_iterations)),
@@ -306,6 +333,10 @@ def train(self, call_back=None):
                         os.remove(os.path.join(self.train_directory, '{}.occs'.format(i)))
                     except FileNotFoundError:
                         pass
+                    try:
+                        os.remove(os.path.join(self.train_directory, '{}.acc'.format(i)))
+                    except FileNotFoundError:
+                        pass
         except Exception as e:
             with open(dirty_path, 'w'):
                 pass