sonusai 0.12.5__tar.gz → 0.12.7__tar.gz

{sonusai-0.12.5 → sonusai-0.12.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sonusai
-Version: 0.12.5
+Version: 0.12.7
 Summary: Framework for building deep neural network models for sound, speech, and voice AI
 Home-page: https://aaware.com
 License: GPL-3.0-only
@@ -29,7 +29,7 @@ Requires-Dist: paho-mqtt (>=1.6.1,<2.0.0)
 Requires-Dist: pandas (>=2.0.3,<3.0.0)
 Requires-Dist: pesq (>=0.0.4,<0.0.5)
 Requires-Dist: protobuf (>=4.24.0,<5.0.0)
-Requires-Dist: pyaaware (>=1.4.18,<2.0.0)
+Requires-Dist: pyaaware (>=1.4.19,<2.0.0)
 Requires-Dist: python-magic (>=0.4.27,<0.5.0)
 Requires-Dist: scikit-learn (>=1.3.0,<2.0.0)
 Requires-Dist: sh (>=1.14.3,<2.0.0)

{sonusai-0.12.5 → sonusai-0.12.7}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "sonusai"
-version = "0.12.5"
+version = "0.12.7"
 description = "Framework for building deep neural network models for sound, speech, and voice AI"
 authors = ["Chris Eddington <chris@aaware.com>", "Jason Calderwood <jason@aaware.com>"]
 maintainers = ["Chris Eddington <chris@aaware.com>", "Jason Calderwood <jason@aaware.com>"]
@@ -31,7 +31,7 @@ paho-mqtt = "^1.6.1"
 pandas = "^2.0.3"
 pesq = "^0.0.4"
 protobuf = "^4.24.0"
-pyaaware = "^1.4.18"
+pyaaware = "^1.4.19"
 python = ">=3.8,<3.11"
 python-magic = "^0.4.27"
 scikit-learn = "^1.3.0"

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/calc_metric_spenh.py

@@ -710,7 +710,7 @@ def _process_mixture(mixid: int) -> Tuple[pd.DataFrame, pd.DataFrame]:
     return all_metrics_table_1, all_metrics_table_2
 
 
-def main():
+def main() -> None:
     from docopt import docopt
 
     import sonusai
@@ -808,7 +808,7 @@ def main():
     # Individual mixtures use pandas print, set precision to 2 decimal places
     # pd.set_option('float_format', '{:.2f}'.format)
     progress = tqdm(total=len(mixids))
-    all_metrics_tables = p_tqdm_map(_process_mixture, mixids, progress=progress)
+    all_metrics_tables = p_tqdm_map(_process_mixture, mixids, progress=progress, chunksize=10)
     progress.close()
 
     all_metrics_table_1 = pd.concat([item[0] for item in all_metrics_tables])

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/data_generator/keras_from_mixdb.py

@@ -65,9 +65,9 @@ class KerasFromMixtureDatabase(Sequence):
         self.stride = self.mixdb.fg.stride
         self.num_bands = self.mixdb.fg.num_bands
         self.num_classes = self.mixdb.num_classes
-        self.mixture_frame_segments = None
-        self.batch_frame_segments = None
-        self.total_batches: Optional[int]
+        self.mixture_frame_segments: Optional[int] = None
+        self.batch_frame_segments: Optional[int] = None
+        self.total_batches: Optional[int] = None
 
         self._initialize_mixtures()
 

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/evaluate.py

@@ -186,7 +186,7 @@ def evaluate(mixdb: MixtureDatabase,
             classdf.round(3).to_csv(join(output_dir, f'class_snr{snri}.csv'))
 
 
-def main():
+def main() -> None:
     from datetime import datetime
     from os import mkdir
     from os.path import join

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/genft.py

@@ -72,7 +72,7 @@ def genft(mixdb: MixtureDatabase,
             results.append(_genft_kernel(mixid))
     else:
         progress = tqdm(total=len(mixids), disable=not show_progress)
-        results = p_tqdm_map(_genft_kernel, mixids, progress=progress)
+        results = p_tqdm_map(_genft_kernel, mixids, progress=progress, chunksize=10)
         progress.close()
 
     return results
@@ -100,7 +100,7 @@ def _genft_kernel(mixid: int) -> GenFTData:
     return GenFTData(feature=feature, truth_f=truth_f, segsnr=segsnr)
 
 
-def main():
+def main() -> None:
     from docopt import docopt
 
     import sonusai

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/genmix.py

@@ -79,7 +79,7 @@ def genmix(mixdb: MixtureDatabase,
             results.append(_genmix_kernel(mixid))
     else:
         progress = tqdm(total=len(mixids), disable=not show_progress)
-        results = p_tqdm_map(_genmix_kernel, mixids, progress=progress)
+        results = p_tqdm_map(_genmix_kernel, mixids, progress=progress, chunksize=10)
         progress.close()
 
     return results
@@ -128,7 +128,7 @@ def _genmix_kernel(mixid: int) -> GenMixData:
                       segsnr_t=segsnr_t)
 
 
-def main():
+def main() -> None:
     from docopt import docopt
 
     import sonusai

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/genmixdb.py

@@ -141,7 +141,6 @@ def genmixdb(location: Location,
              show_progress: bool = False,
              test_mode: bool = False,
              use_cache: bool = True) -> MixtureDatabase:
-    import itertools
     from random import seed
 
     import yaml
@@ -155,7 +154,6 @@ def genmixdb(location: Location,
     from sonusai.mixture import SAMPLE_RATE
     from sonusai.mixture import TruthSettings
     from sonusai.mixture import balance_targets
-    from sonusai.mixture import estimate_augmented_length_from_audio
     from sonusai.mixture import generate_mixtures
     from sonusai.mixture import get_augmentation_indices_for_mixup
     from sonusai.mixture import get_augmentations
@@ -286,19 +284,11 @@ def genmixdb(location: Location,
 
     augmented_targets = balance_targets(mixdb, augmented_targets)
 
-    augmented_noise_iter = list(itertools.product(*[range(len(mixdb.noises)), range(len(mixdb.noise_augmentations))]))
-    noise_audio_samples = sum([mixdb.augmented_noise_length(nfi, nai) for nfi, nai in augmented_noise_iter])
-
-    if logging:
-        logger.info('Generating mixtures first pass')
-    used_noise_files, used_noise_samples = generate_mixtures(mixdb=mixdb,
-                                                             augmented_targets=augmented_targets,
-                                                             noise_files=noise_files,
-                                                             noise_augmentations=noise_augmentations,
-                                                             mixups=mixups,
-                                                             show_progress=show_progress)
-
     total_noise_files = len(mixdb.noises) * len(mixdb.noise_augmentations)
+    aug_noise_audio_samples = mixdb.augmented_noise_samples
+
+    total_target_files = len(augmented_targets)
+    aug_target_audio_samples = mixdb.augmented_target_samples
 
     if logging:
         raw_target_audio_samples = sum([targets.samples for targets in mixdb.targets])
@@ -312,24 +302,19 @@ def genmixdb(location: Location,
                     f'{human_readable_size(raw_noise_audio_duration * SAMPLE_RATE * SAMPLE_BYTES, 1)}, '
                     f'{seconds_to_hms(seconds=raw_noise_audio_duration)}')
 
-        augmented_noise_audio_samples = float(sum([mixdb.augmented_noise_length(f, a) for f, a in
-                                                   zip(range(len(mixdb.noises)),
-                                                       range(len(mixdb.noise_augmentations)))]))
-        augmented_target_audio_samples = 0
-        for augmented_target in augmented_targets:
-            augmented_target_audio_samples += estimate_augmented_length_from_audio(
-                audio=mixdb.raw_target_audio(augmented_target.target_file_index),
-                augmentation=mixdb.target_augmentations[augmented_target.target_augmentation_index],
-                length_common_denominator=mixdb.feature_step_samples)
-
         logger.info('')
-        logger.info(f'Augmented target audio: {len(augmented_targets)} files, '
-                    f'{human_readable_size(augmented_target_audio_samples * SAMPLE_BYTES, 1)}, '
-                    f'{seconds_to_hms(seconds=augmented_target_audio_samples / SAMPLE_RATE)}')
+        logger.info(f'Augmented target audio: {total_target_files} files, '
+                    f'{human_readable_size(aug_target_audio_samples * SAMPLE_BYTES, 1)}, '
+                    f'{seconds_to_hms(seconds=aug_target_audio_samples / SAMPLE_RATE)}')
         logger.info(f'Augmented noise audio: {total_noise_files} files, '
-                    f'{human_readable_size(augmented_noise_audio_samples * SAMPLE_BYTES, 1)}, '
-                    f'{seconds_to_hms(seconds=augmented_noise_audio_samples / SAMPLE_RATE)}')
+                    f'{human_readable_size(aug_noise_audio_samples * SAMPLE_BYTES, 1)}, '
+                    f'{seconds_to_hms(seconds=aug_noise_audio_samples / SAMPLE_RATE)}')
 
+    used_noise_files, used_noise_samples = generate_mixtures(mixdb=mixdb,
+                                                             augmented_targets=augmented_targets,
+                                                             noise_files=noise_files,
+                                                             noise_augmentations=noise_augmentations,
+                                                             mixups=mixups)
     total_mixtures = len(mixdb.mixtures)
     if logging:
         logger.info('')
@@ -358,9 +343,9 @@ def genmixdb(location: Location,
     MP_GLOBAL.save_segsnr = save_segsnr
 
     if logging:
-        logger.info('Generating mixtures second pass')
+        logger.info('Generating mixtures')
     progress = tqdm(total=total_mixtures, disable=not show_progress)
-    mixdb.mixtures = p_tqdm_map(_process_mixture, range(total_mixtures), progress=progress)
+    mixdb.mixtures = p_tqdm_map(_process_mixture, range(total_mixtures), progress=progress, chunksize=10)
     progress.close()
 
     total_samples = mixdb.total_samples()
@@ -374,7 +359,7 @@ def genmixdb(location: Location,
                                stride=mixdb.fg.stride,
                                desc='Actual')
         noise_files_percent = (float(used_noise_files) / float(total_noise_files)) * 100
-        noise_samples_percent = (float(used_noise_samples) / float(noise_audio_samples)) * 100
+        noise_samples_percent = (float(used_noise_samples) / float(aug_noise_audio_samples)) * 100
         logger.info('')
         logger.info(f'Used {noise_files_percent:,.0f}% of augmented noise files')
         logger.info(f'Used {noise_samples_percent:,.0f}% of augmented noise audio')
@@ -431,7 +416,7 @@ def _process_mixture(mixid: int) -> MRecord:
     return mrecord
 
 
-def main():
+def main() -> None:
     from docopt import docopt
 
     import sonusai

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/gentcst.py

@@ -579,7 +579,7 @@ def report_leaf_fold_data_usage(all_files: List[FileInfo], use_files: List[FileI
         logger.warning('')
 
 
-def main():
+def main() -> None:
     from docopt import docopt
 
     import sonusai

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/keras_onnx.py

@@ -30,7 +30,7 @@ Results are written into subdirectory <MODEL>-<TIMESTAMP> unless OUTPUT is speci
 from sonusai import logger
 
 
-def main():
+def main() -> None:
     from docopt import docopt
 
     import sonusai

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/keras_predict.py

@@ -40,7 +40,7 @@ from sonusai.mixture import Feature
 from sonusai.mixture import Predict
 
 
-def main():
+def main() -> None:
     from docopt import docopt
 
     import sonusai

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/keras_train.py

@@ -80,7 +80,7 @@ class SonusAIModelCheckpoint(tf.keras.callbacks.ModelCheckpoint):
                 f.attrs['sonusai_num_classes'] = str(self.num_classes)
 
 
-def main():
+def main() -> None:
     from docopt import docopt
 
     import sonusai
@@ -285,9 +285,9 @@ def main():
     # Find checkpoint file and load weights for prediction and model save
     checkpoint_name = None
     for path, dirs, files in walk(output_dir):
-        for f in files:
-            if "ckpt" in f:
-                checkpoint_name = f
+        for file in files:
+            if "ckpt" in file:
+                checkpoint_name = file
 
     if checkpoint_name is not None:
         logger.info('Using best checkpoint for prediction and model exports')

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/lsdb.py

@@ -109,7 +109,7 @@ def lsdb(mixdb: MixtureDatabase,
             logger.info(f'  {snr_mean[c]:8.2f}  {snr_std[c]:8.2f}  {snr_db_mean[c]:8.2f}  {snr_db_std[c]:8.2f}')
 
 
-def main():
+def main() -> None:
     from docopt import docopt
 
     import sonusai

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/main.py

@@ -4,7 +4,6 @@ usage: sonusai [--version] [--help] <command> [<args>...]
 
 The sonusai commands are:
    calc_metric_spenh            Run speech enhancement and analysis
-   calc_metric_spenh_targetf    Run speech enhancement and analysis for targetf truth (deprecated)
    evaluate                     Evaluate model performance
    genft                        Generate feature and truth data
    genmix                       Generate mixture and truth data
@@ -14,6 +13,7 @@ The sonusai commands are:
    keras_train                  Train a model using Keras
    keras_onnx                   Convert a trained Keras model to ONNX
    lsdb                         List information about a mixture database
+   mkmanifest                   Make ASR manifest JSON file
    mkwav                        Make WAV files from a mixture database
    onnx_predict                 Run ONNX predict on a trained model
    plot                         Plot mixture data
@@ -30,18 +30,14 @@ for more information on a specific command.
 from sonusai import logger
 
 
-def main():
-    from subprocess import call
-
+def main() -> None:
     from docopt import docopt
 
     import sonusai
-    from sonusai import SonusAIError
     from sonusai.utils import trim_docstring
 
     commands = (
         'calc_metric_spenh',
-        'calc_metric_spenh_targetf',
         'evaluate',
         'genft',
         'genmix',
@@ -51,6 +47,7 @@ def main():
         'keras_train',
         'keras_onnx',
         'lsdb',
+        'mkmanifest',
         'mkwav',
         'onnx_predict',
         'plot',
@@ -66,6 +63,11 @@ def main():
     command = args['<command>']
     argv = args['<args>']
 
+    from subprocess import call
+
+    import sonusai
+    from sonusai import SonusAIError
+
     if command == 'help':
         if not argv:
             exit(call(['sonusai', '-h']))

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/mixture/__init__.py

@@ -1,5 +1,5 @@
 # SonusAI mixture utilities
-from sonusai.mixture.audio import AawareSoxTransformer
+from sonusai.mixture.audio import Transformer
 from sonusai.mixture.audio import calculate_audio_from_transform
 from sonusai.mixture.audio import calculate_transform_from_audio
 from sonusai.mixture.audio import get_duration

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/mixture/audio.py

@@ -97,7 +97,6 @@ def read_audio(name: Location) -> AudioT:
     from typing import Any
 
     import numpy as np
-    import sox
 
     from sonusai import SonusAIError
     from sonusai.mixture import BIT_DEPTH
@@ -196,10 +195,16 @@ def get_duration(audio: AudioT) -> float:
     return len(audio) / SAMPLE_RATE
 
 
-class AawareSoxTransformer(sox.Transformer):
+class Transformer(sox.Transformer):
+    """Override certain sox.Transformer methods
+    """
+
     def fir(self, coefficients):
         """Use SoX’s FFT convolution engine with given FIR filter coefficients.
 
+        The SonusAI override allows coefficients to be either a list of numbers
+        or a string containing a text file with the coefficients.
+
         Parameters
         ----------
         coefficients : list or str
@@ -225,6 +230,69 @@ class AawareSoxTransformer(sox.Transformer):
 
         return self
 
+    def tempo(self, factor, audio_type=None, quick=False):
+        """Time stretch audio without changing pitch.
+
+        This effect uses the WSOLA algorithm. The audio is chopped up into
+        segments which are then shifted in the time domain and overlapped
+        (cross-faded) at points where their waveforms are most similar as
+        determined by measurement of least squares.
+
+        The SonusAI override does not generate a warning for small factors.
+        The sox.Transformer's implementation of stretch does not invert
+        the factor even though it says that it does; this invalidates the
+        factor size check and produces the wrong result.
+
+        Parameters
+        ----------
+        factor : float
+            The ratio of new tempo to the old tempo.
+            For ex. 1.1 speeds up the tempo by 10%; 0.9 slows it down by 10%.
+        audio_type : str
+            Type of audio, which optimizes algorithm parameters. One of:
+             * m : Music,
+             * s : Speech,
+             * l : Linear (useful when factor is close to 1),
+        quick : bool, default=False
+            If True, this effect will run faster but with lower sound quality.
+
+        See Also
+        --------
+        stretch, speed, pitch
+
+        """
+        from sox.core import is_number
+
+        from sonusai import SonusAIError
+        from sonusai import logger
+
+        if not is_number(factor) or factor <= 0:
+            raise SonusAIError('factor must be a positive number')
+
+        if factor < 0.5 or factor > 2:
+            logger.warning('Using an extreme time stretching factor. Quality of results will be poor')
+
+        if audio_type not in [None, 'm', 's', 'l']:
+            raise SonusAIError("audio_type must be one of None, 'm', 's', or 'l'.")
+
+        if not isinstance(quick, bool):
+            raise SonusAIError('quick must be a boolean')
+
+        effect_args = ['tempo']
+
+        if quick:
+            effect_args.append('-q')
+
+        if audio_type is not None:
+            effect_args.append('-{}'.format(audio_type))
+
+        effect_args.append('{:f}'.format(factor))
+
+        self.effects.extend(effect_args)
+        self.effects_log.append('tempo')
+
+        return self
+
 
 def validate_input_file(input_filepath: str) -> None:
     from os.path import exists
@@ -239,4 +307,4 @@ def validate_input_file(input_filepath: str) -> None:
 
     ext = splitext(input_filepath)[1][1:].lower()
     if ext not in VALID_FORMATS:
-        raise SonusAIError(f'This installation of Sox cannot process .{ext} files')
+        raise SonusAIError(f'This installation of SoX cannot process .{ext} files')

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/mixture/augmentation.py

@@ -173,17 +173,17 @@ def apply_augmentation(audio: AudioT, augmentation: Augmentation, length_common_
     :param length_common_denominator: Pad resulting audio to be a multiple of this
     :return: Augmented audio
     """
-    import sox
 
     from sonusai import SonusAIError
     from sonusai.mixture import BIT_DEPTH
     from sonusai.mixture import CHANNEL_COUNT
     from sonusai.mixture import ENCODING
     from sonusai.mixture import SAMPLE_RATE
+    from sonusai.mixture import Transformer
 
     try:
         # Apply augmentations
-        tfm = sox.Transformer()
+        tfm = Transformer()
         tfm.set_input_format(rate=SAMPLE_RATE, bits=BIT_DEPTH, channels=CHANNEL_COUNT, encoding=ENCODING)
         tfm.set_output_format(rate=SAMPLE_RATE, bits=BIT_DEPTH, channels=CHANNEL_COUNT, encoding=ENCODING)
 
@@ -201,11 +201,7 @@ def apply_augmentation(audio: AudioT, augmentation: Augmentation, length_common_
             tfm.pitch(n_semitones=float(augmentation.pitch) / 100)
 
         if augmentation.tempo is not None:
-            factor = float(augmentation.tempo)
-            if abs(factor - 1.0) <= 0.1:
-                tfm.stretch(factor=factor)
-            else:
-                tfm.tempo(factor=factor, audio_type='s')
+            tfm.tempo(factor=float(augmentation.tempo), audio_type='s')
 
         if augmentation.eq1 is not None:
             tfm.equalizer(frequency=augmentation.eq1[0], width_q=augmentation.eq1[1],
@@ -240,11 +236,10 @@ def apply_ir(audio: AudioT, ir: ImpulseResponseData) -> AudioT:
     :return: Augmented audio
     """
     import numpy as np
-    import sox
 
     from sonusai import SonusAIError
-    from sonusai.mixture import AawareSoxTransformer
     from sonusai.mixture import SAMPLE_RATE
+    from sonusai.mixture import Transformer
     from sonusai.utils import linear_to_db
 
     max_abs_audio = max(abs(audio))
@@ -257,7 +252,7 @@ def apply_ir(audio: AudioT, ir: ImpulseResponseData) -> AudioT:
     max_db = linear_to_db(max_abs_audio)
 
     # Convert audio to IR sample rate and normalize to -20 dBFS to avoid clipping when applying IR
-    tfm = sox.Transformer()
+    tfm = Transformer()
     tfm.set_output_format(rate=ir.sample_rate)
     tfm.norm(db_level=-20)
     audio_out = tfm.build_array(input_array=audio, sample_rate_in=SAMPLE_RATE)
@@ -267,7 +262,7 @@ def apply_ir(audio: AudioT, ir: ImpulseResponseData) -> AudioT:
     audio_out = np.pad(array=audio_out, pad_width=(pad, pad))
 
     # Apply IR and convert back to global sample rate
-    tfm = AawareSoxTransformer()
+    tfm = Transformer()
     tfm.set_output_format(rate=SAMPLE_RATE)
     tfm.fir(coefficients=ir.coefficients_file)
     try:
@@ -276,7 +271,7 @@ def apply_ir(audio: AudioT, ir: ImpulseResponseData) -> AudioT:
         raise SonusAIError(f'Error applying IR: {e}')
 
     # Reset level to previous max value
-    tfm = sox.Transformer()
+    tfm = Transformer()
     tfm.norm(db_level=max_db)
     audio_out = tfm.build_array(input_array=audio_out, sample_rate_in=SAMPLE_RATE)
 
@@ -296,8 +291,10 @@ def estimate_augmented_length_from_length(length: int,
     :param length_common_denominator: Pad resulting audio to be a multiple of this
     :return: Estimated length of augmented audio
     """
+    import numpy as np
+
     if augmentation.tempo is not None:
-        length = int(length // float(augmentation.tempo))
+        length = int(np.round(length / float(augmentation.tempo)))
 
     length += get_pad_length(length, length_common_denominator)
 

{sonusai-0.12.5 → sonusai-0.12.7}/sonusai/mixture/config.py

@@ -250,7 +250,7 @@ def get_target_files(config: dict, show_progress: bool = False) -> TargetFiles:
                                              for target in config['targets']]))
 
     progress = tqdm(total=len(target_files), disable=not show_progress)
-    target_files = p_tqdm_map(_get_samples, target_files, progress=progress)
+    target_files = p_tqdm_map(_get_samples, target_files, progress=progress, chunksize=10)
     progress.close()
 
     max_class = get_max_class(config['num_classes'], config['truth_mode'] == 'mutex')
@@ -394,7 +394,7 @@ def get_noise_files(config: dict, show_progress: bool = False) -> NoiseFiles:
     noise_files = list(chain.from_iterable([_append_noise_files(noise_file=noise) for noise in config['noises']]))
 
     progress = tqdm(total=len(noise_files), disable=not show_progress)
-    noise_files = p_tqdm_map(_get_samples, noise_files, progress=progress)
+    noise_files = p_tqdm_map(_get_samples, noise_files, progress=progress, chunksize=10)
     progress.close()
 
     return dataclass_from_dict(NoiseFiles, noise_files)