Ignore differences in SR that can be handled in Kaldi feature generat…

…ion (#276)

Fixes for gating on sample rates, see changelog for more details

.deepsource.toml

@@ -8,3 +8,4 @@ enabled = true
 
   [analyzers.meta]
   runtime_version = "3.x.x"
+  max_line_length = 120

docs/source/changelog.rst

@@ -6,6 +6,20 @@
 Changelog
 =========
 
+2.0.0a12
+--------
+
+- Updated how sample rates are handled. MFA now generates features between 80 Hz and 7800 Hz and allows downsampling and
+  upsampling, so there will be no more errors or warnings about unsupported sample rates or speakers with different sample
+  rates
+- Fixed a bug where some options for generating MFCCs weren't properly getting picked up (e.g., snip-edges)
+- (EXPERIMENTAL) Added better support for varying frame shift. In :code:`mfa align`, you can now add a flag of :code:`--frame_shift 1` to align
+  with millisecond shifts between frames.  Please note this is more on the experimental side, as it increases computational
+  time significantly and I don't know fully the correct options to use for :code:`self_loop_scale`, :code:`transition_scale`,
+  and :code:`acoustic_scale` to generate good alignments.
+- Fixed a bug in G2P training with relative paths for output model
+- Cleaned up validator output
+
 2.0.0a11
 --------
 

montreal_forced_aligner/__init__.py

@@ -1,6 +1,6 @@
 __ver_major__ = 2
 __ver_minor__ = 0
-__ver_patch__ = '0a11'
+__ver_patch__ = '0a13'
 __version__ = "{}.{}.{}".format(__ver_major__, __ver_minor__, __ver_patch__)
 
 __all__ = ['aligner', 'command_line', 'models', 'corpus', 'config', 'dictionary', 'exceptions',

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/g2p.py

@@ -5,7 +5,8 @@
 from textgrid import TextGrid, IntervalTier
 
 from montreal_forced_aligner.multiprocessing.corpus import parse_transcription
-from montreal_forced_aligner.corpus.align_corpus import load_text, AlignableCorpus
+from montreal_forced_aligner.corpus.align_corpus import AlignableCorpus
+from montreal_forced_aligner.helper import load_text
 
 from montreal_forced_aligner.g2p.generator import PyniniDictionaryGenerator as Generator
 from montreal_forced_aligner.models import G2PModel

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
@@ -38,7 +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:
             conf = yaml.load(f, Loader=yaml.SafeLoader)

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

@@ -4,9 +4,9 @@
 import re
 from collections import Counter
 from textgrid import TextGrid, IntervalTier
-from ..helper import load_text, output_mapping, save_groups, filter_scp, load_scp
+from ..helper import output_mapping, save_groups, filter_scp, load_scp
 
-from ..exceptions import SampleRateError, CorpusError, WavReadError, SampleRateMismatchError, \
+from ..exceptions import CorpusError, WavReadError,  \
     BitDepthError, TextParseError, TextGridParseError
 
 from .base import BaseCorpus, find_exts
@@ -65,11 +65,14 @@ def __init__(self, directory, output_directory,
 
         self.loaded_from_temp = self._load_from_temp()
         if not self.loaded_from_temp:
-            print('loading from source')
             if self.use_mp:
+                self.logger.debug('Loading from source with multiprocessing')
                 self._load_from_source_mp()
             else:
+                self.logger.debug('Loading from source without multiprocessing')
                 self._load_from_source()
+        else:
+            self.logger.debug('Successfully loaded from temporary files')
         self.check_warnings()
         self.find_best_groupings()
 
@@ -129,7 +132,6 @@ def _load_from_temp(self):
                 new_w = re.split(r"[-']", w)
                 self.word_counts.update(new_w + [w])
             self.text_mapping[utt] = ' '.join(text)
-        self.sample_rates = {int(k): v if isinstance(v, list) else [v] for k,v in load_scp(sr_path).items()}
         self.utt_wav_mapping = load_scp(wav_path)
         self.wav_info = load_scp(wav_info_path, float)
         self.utt_text_file_mapping = load_scp(text_file_path)
@@ -189,7 +191,7 @@ def _load_from_source_mp(self):
                 else:
                     self.no_transcription_files.append(wav_path)
                     continue
-                job_queue.put((file_name, wav_path, transcription_path, relative_path, self.speaker_characters, self.temp_directory))
+                job_queue.put((file_name, wav_path, transcription_path, relative_path, self.speaker_characters))
         job_queue.join()
         stopped.stop()
         for p in procs:
@@ -204,7 +206,6 @@ def _load_from_source_mp(self):
                 utt_name = info['utt_name']
                 speaker_name = info['speaker_name']
                 wav_info = info['wav_info']
-                sr = wav_info['sample_rate']
                 if utt_name in self.utt_wav_mapping:
                     ind = 0
                     fixed_utt_name = utt_name
@@ -218,22 +219,19 @@ def _load_from_source_mp(self):
                 for w in words:
                     new_w = re.split(r"[-']", w)
                     self.word_counts.update(new_w + [w])
+                self.wav_files.append(file_name)
                 self.text_mapping[utt_name] = ' '.join(words)
                 self.utt_text_file_mapping[utt_name] = info['text_file']
                 self.speak_utt_mapping[speaker_name].append(utt_name)
                 self.utt_wav_mapping[utt_name] = info['wav_path']
-                self.sample_rates[sr].add(speaker_name)
                 self.utt_speak_mapping[utt_name] = speaker_name
                 self.file_directory_mapping[utt_name] = info['relative_path']
                 self.lab_count += 1
             else:
                 wav_info = info['wav_info']
-                sr = wav_info['sample_rate']
                 file_name = info['recording_name']
                 self.wav_files.append(file_name)
                 self.speaker_ordering[file_name] = info['speaker_ordering']
-                for s in info['speaker_ordering']:
-                    self.sample_rates[sr].add(s)
                 self.segments.update(info['segments'])
                 self.utt_wav_mapping.update(info['utt_wav_mapping'])
                 self.file_utt_mapping.update(info['file_utt_mapping'])
@@ -255,7 +253,7 @@ def _load_from_source_mp(self):
                     self.file_directory_mapping[fn] = info['relative_path']
                 self.tg_count += 1
             self.wav_info[file_name] = [wav_info['num_channels'], wav_info['sample_rate'], wav_info['duration']]
-        for k in ['wav_read_errors', 'unsupported_sample_rate', 'unsupported_bit_depths',
+        for k in ['wav_read_errors', 'unsupported_bit_depths',
                   'decode_error_files', 'textgrid_read_errors']:
             if hasattr(self, k):
                 if k in return_dict:
@@ -280,7 +278,6 @@ def _load_from_source(self):
                         utt_name = info['utt_name']
                         speaker_name = info['speaker_name']
                         wav_info = info['wav_info']
-                        sr = wav_info['sample_rate']
                         if utt_name in self.utt_wav_mapping:
                             ind = 0
                             fixed_utt_name = utt_name
@@ -298,14 +295,11 @@ def _load_from_source(self):
                         self.utt_text_file_mapping[utt_name] = lab_path
                         self.speak_utt_mapping[speaker_name].append(utt_name)
                         self.utt_wav_mapping[utt_name] = wav_path
-                        self.sample_rates[sr].add(speaker_name)
                         self.utt_speak_mapping[utt_name] = speaker_name
                         self.file_directory_mapping[utt_name] = relative_path
                         self.lab_count += 1
                     except WavReadError:
                         self.wav_read_errors.append(wav_path)
-                    except SampleRateError:
-                        self.unsupported_sample_rate.append(wav_path)
                     except BitDepthError:
                         self.unsupported_bit_depths.append(wav_path)
                     except TextParseError:
@@ -316,13 +310,10 @@ def _load_from_source(self):
                     tg_path = os.path.join(root, tg_name)
                     try:
                         info = parse_textgrid_file(file_name, wav_path, tg_path, relative_path,
-                                                   self.speaker_characters, self.temp_directory)
+                                                   self.speaker_characters)
                         wav_info = info['wav_info']
-                        sr = wav_info['sample_rate']
                         self.wav_files.append(file_name)
                         self.speaker_ordering[file_name] = info['speaker_ordering']
-                        for s in info['speaker_ordering']:
-                            self.sample_rates[sr].add(s)
                         self.segments.update(info['segments'])
                         self.utt_wav_mapping.update(info['utt_wav_mapping'])
                         self.utt_text_file_mapping.update(info['utt_text_file_mapping'])
@@ -343,8 +334,6 @@ def _load_from_source(self):
                         self.tg_count += 1
                     except WavReadError:
                         self.wav_read_errors.append(wav_path)
-                    except SampleRateError:
-                        self.unsupported_sample_rate.append(wav_path)
                     except BitDepthError:
                         self.unsupported_bit_depths.append(wav_path)
                     except TextGridParseError as e:
@@ -358,30 +347,7 @@ def _load_from_source(self):
 
     def check_warnings(self):
         self.issues_check = self.ignored_utterances or self.no_transcription_files or \
-                            self.textgrid_read_errors or self.unsupported_sample_rate or self.decode_error_files
-
-        bad_speakers = []
-        for speaker in self.speak_utt_mapping.keys():
-            count = 0
-            for k, v in self.sample_rates.items():
-                if speaker in v:
-                    count += 1
-            if count > 1:
-                bad_speakers.append(speaker)
-        if bad_speakers:
-            msg = 'The following speakers had multiple speaking rates: {}. ' \
-                  'Please make sure that each speaker has a consistent sampling rate.'.format(', '.join(bad_speakers))
-            self.logger.error(msg)
-            raise (SampleRateMismatchError(msg))
-
-        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)
+                            self.textgrid_read_errors or self.decode_error_files
 
     def save_text_file(self, file_name):
         if self.segments:

montreal_forced_aligner/corpus/base.py

@@ -137,6 +137,10 @@ def __init__(self, directory, output_directory,
         if os.path.exists(feat_path):
             self.feat_mapping = load_scp(feat_path)
 
+    @property
+    def speakers(self):
+        return sorted(self.speak_utt_mapping.keys())
+
     def add_utterance(self, utterance, speaker, file, text, wav_file=None, seg=None):
         if seg is not None:
             self.segments[utterance] = seg
@@ -171,35 +175,15 @@ def delete_utterance(self, utterance):
             del self.utt_wav_mapping[utterance]
 
     def find_best_groupings(self):
-        num_sample_rates = len(self.sample_rates.keys())
-        jobs_per_sample_rate = {x: 1 for x in self.sample_rates.keys()}
-        remaining_jobs = self.num_jobs - num_sample_rates
-        while remaining_jobs > 0:
-            min_num = min(jobs_per_sample_rate.values())
-            addable = sorted([k for k, v in jobs_per_sample_rate.items() if v == min_num],
-                             key=lambda x: -1 * len(self.sample_rates[x]))
-            jobs_per_sample_rate[addable[0]] += 1
-            remaining_jobs -= 1
-        self.speaker_groups = []
-        self.frequency_configs = []
-        job_num = 0
-        for k, v in jobs_per_sample_rate.items():
-            speakers = sorted(self.sample_rates[k])
-
-            groups = [[] for _ in range(v)]
-
-            configs = [(job_num + x, {'sample-frequency': k, 'low-freq': 20, 'high-freq': 7800}) for x in range(v)]
-            ind = 0
-            while speakers:
-                s = speakers.pop(0)
-                groups[ind].append(s)
-                ind += 1
-                if ind >= v:
-                    ind = 0
-
-            job_num += v
-            self.speaker_groups.extend(groups)
-            self.frequency_configs.extend(configs)
+        if len(self.speakers) < self.num_jobs:
+            self.num_jobs = len(self.speakers)
+        self.speaker_groups = [[] for _ in range(self.num_jobs)]
+        job_ind = 0
+        for s in self.speakers:
+            self.speaker_groups[job_ind].append(s)
+            job_ind += 1
+            if job_ind == self.num_jobs:
+                job_ind = 0
         self.groups = []
         for x in self.speaker_groups:
             g = []