sonusai 0.15.6__py3-none-any.whl → 0.15.9__py3-none-any.whl

sonusai/audiofe.py

@@ -0,0 +1,293 @@
+"""sonusai audiofe
+
+usage: audiofe [-hvds] [--version] [-i INPUT] [-l LENGTH] [-m MODEL] [-k CKPT] [-a ASR] [-w WMODEL]
+
+options:
+    -h, --help
+    -v, --verbose                   Be verbose.
+    -d, --debug                     Write debug data to H5 file.
+    -s, --show                      Show a list of available audio inputs.
+    -i INPUT, --input INPUT         Input audio.
+    -l LENGTH, --length LENGTH      Length of audio in seconds. [default: -1].
+    -m MODEL, --model MODEL         PL model .py file path.
+    -k CKPT, --checkpoint CKPT      PL checkpoint file with weights.
+    -a ASR, --asr ASR               ASR method to use.
+    -w WMODEL, --whisper WMODEL     Whisper model used in aixplain_whisper and whisper methods. [default: tiny].
+
+Aaware SonusAI Audio Front End.
+
+Capture LENGTH seconds of audio from INPUT. If LENGTH is < 0, then capture until key is pressed. If INPUT is a valid
+audio file name, then use the audio data from the specified file. In this case, if LENGTH is < 0, process entire file;
+otherwise, process min(length(INPUT), LENGTH) seconds of audio from INPUT. Audio is saved to
+audiofe_capture_<TIMESTAMP>.wav.
+
+If a model is specified, run prediction on audio data from this model. Then compute the inverse transform of the
+prediction result and save to audiofe_predict_<TIMESTAMP>.wav.
+
+If an ASR is specified, run ASR on the captured audio and print the results. In addition, if a model was also specified,
+run ASR on the predict audio and print the results.
+
+If the debug option is enabled, write capture audio, feature, reconstruct audio, predict, and predict audio to
+audiofe_<TIMESTAMP>.h5.
+
+"""
+from os.path import exists
+from select import select
+from sys import stdin
+from typing import Any
+
+import h5py
+import numpy as np
+import pyaudio
+import torch
+from docopt import docopt
+from docopt import printable_usage
+
+import sonusai
+from sonusai import create_file_handler
+from sonusai import initial_log_messages
+from sonusai import logger
+from sonusai import update_console_handler
+from sonusai.mixture import AudioT
+from sonusai.mixture import CHANNEL_COUNT
+from sonusai.mixture import SAMPLE_RATE
+from sonusai.mixture import get_audio_from_feature
+from sonusai.mixture import get_feature_from_audio
+from sonusai.mixture import read_audio
+from sonusai.utils import calc_asr
+from sonusai.utils import create_timestamp
+from sonusai.utils import get_input_device_index_by_name
+from sonusai.utils import get_input_devices
+from sonusai.utils import import_keras_model
+from sonusai.utils import trim_docstring
+from sonusai.utils import write_wav
+
+
+def main() -> None:
+    args = docopt(trim_docstring(__doc__), version=sonusai.__version__, options_first=True)
+    ts = create_timestamp()
+
+    verbose = args['--verbose']
+    length = float(args['--length'])
+    input_name = args['--input']
+    model_name = args['--model']
+    ckpt_name = args['--checkpoint']
+    asr_name = args['--asr']
+    whisper_name = args['--whisper']
+    debug = args['--debug']
+    show = args['--show']
+
+    capture_name = f'audiofe_capture_{ts}.wav'
+    predict_name = f'audiofe_predict_{ts}.wav'
+    h5_name = f'audiofe_{ts}.h5'
+
+    if model_name is not None and ckpt_name is None:
+        print(printable_usage(trim_docstring(__doc__)))
+        exit(1)
+
+    # Setup logging file
+    create_file_handler('audiofe.log')
+    update_console_handler(verbose)
+    initial_log_messages('audiofe')
+
+    if show:
+        logger.info('List of available audio inputs:')
+        logger.info('')
+        p = pyaudio.PyAudio()
+        for name in get_input_devices(p):
+            logger.info(f'{name}')
+        logger.info('')
+        p.terminate()
+        return
+
+    if input_name is not None and exists(input_name):
+        capture_audio = get_frames_from_file(input_name, length)
+    else:
+        try:
+            capture_audio = get_frames_from_device(input_name, length)
+        except ValueError as e:
+            logger.exception(e)
+            return
+
+    write_wav(capture_name, capture_audio, SAMPLE_RATE)
+    logger.info('')
+    logger.info(f'Wrote capture audio with shape {capture_audio.shape} to {capture_name}')
+    if debug:
+        with h5py.File(h5_name, 'a') as f:
+            if 'capture_audio' in f:
+                del f['capture_audio']
+            f.create_dataset('capture_audio', data=capture_audio)
+        logger.info(f'Wrote capture audio with shape {capture_audio.shape} to {h5_name}')
+
+    if asr_name is not None:
+        capture_asr = calc_asr(capture_audio, engine=asr_name, whisper_model_name=whisper_name).text
+        logger.info(f'Capture audio ASR: {capture_asr}')
+
+    if model_name is not None:
+        model = load_model(model_name=model_name, ckpt_name=ckpt_name)
+
+        feature = get_feature_from_audio(audio=capture_audio, feature_mode=model.hparams.feature)
+        if debug:
+            with h5py.File(h5_name, 'a') as f:
+                if 'feature' in f:
+                    del f['feature']
+                f.create_dataset('feature', data=feature)
+            logger.info(f'Wrote feature with shape {feature.shape} to {h5_name}')
+
+        # if debug:
+        #     reconstruct_name = f'audiofe_reconstruct_{ts}.wav'
+        #     reconstruct_audio = get_audio_from_feature(feature=feature, feature_mode=model.hparams.feature)
+        #     samples = min(len(capture_audio), len(reconstruct_audio))
+        #     max_err = np.max(np.abs(capture_audio[:samples] - reconstruct_audio[:samples]))
+        #     logger.info(f'Maximum error between capture and reconstruct: {max_err}')
+        #     write_wav(reconstruct_name, reconstruct_audio, SAMPLE_RATE)
+        #     logger.info(f'Wrote reconstruct audio with shape {reconstruct_audio.shape} to {reconstruct_name}')
+        #     with h5py.File(h5_name, 'a') as f:
+        #         if 'reconstruct_audio' in f:
+        #             del f['reconstruct_audio']
+        #         f.create_dataset('reconstruct_audio', data=reconstruct_audio)
+        #     logger.info(f'Wrote reconstruct audio with shape {reconstruct_audio.shape} to {h5_name}')
+
+        with torch.no_grad():
+            predict = model(torch.tensor(feature))
+        if debug:
+            with h5py.File(h5_name, 'a') as f:
+                if 'predict' in f:
+                    del f['predict']
+                f.create_dataset('predict', data=predict)
+            logger.info(f'Wrote predict with shape {predict.shape} to {h5_name}')
+
+        predict_audio = get_audio_from_feature(feature=predict.numpy(), feature_mode=model.hparams.feature)
+        write_wav(predict_name, predict_audio, SAMPLE_RATE)
+        logger.info(f'Wrote predict audio with shape {predict_audio.shape} to {predict_name}')
+        if debug:
+            with h5py.File(h5_name, 'a') as f:
+                if 'predict_audio' in f:
+                    del f['predict_audio']
+                f.create_dataset('predict_audio', data=predict_audio)
+            logger.info(f'Wrote predict audio with shape {predict_audio.shape} to {h5_name}')
+
+        if asr_name is not None:
+            predict_asr = calc_asr(predict_audio, engine=asr_name, whisper_model_name=whisper_name).text
+            logger.info(f'Predict audio ASR: {predict_asr}')
+
+
+def load_model(model_name: str, ckpt_name: str) -> Any:
+    batch_size = 1
+    timesteps = 0
+
+    # Load checkpoint first to get hparams if available
+    try:
+        checkpoint = torch.load(ckpt_name, map_location=lambda storage, loc: storage)
+    except Exception as e:
+        logger.exception(f'Error: could not load checkpoint from {ckpt_name}: {e}')
+        raise SystemExit(1)
+
+    # Import model definition file
+    logger.info(f'Importing {model_name}')
+    litemodule = import_keras_model(model_name)
+
+    if 'hyper_parameters' in checkpoint:
+        logger.info(f'Found checkpoint file with hyper-parameters')
+        hparams = checkpoint['hyper_parameters']
+        if hparams['batch_size'] != batch_size:
+            logger.info(
+                f'Overriding model default batch_size of {hparams["batch_size"]} with batch_size of {batch_size}')
+            hparams["batch_size"] = batch_size
+
+        if hparams['timesteps'] != 0 and timesteps == 0:
+            timesteps = hparams['timesteps']
+            logger.warning(f'Using model default timesteps of {timesteps}')
+
+        logger.info(f'Building model with {len(hparams)} total hparams')
+        try:
+            model = litemodule.MyHyperModel(**hparams)
+        except Exception as e:
+            logger.exception(f'Error: model build (MyHyperModel) in {model_name} failed: {e}')
+            raise SystemExit(1)
+    else:
+        logger.info(f'Found checkpoint file with no hyper-parameters')
+        logger.info(f'Building model with defaults')
+        try:
+            tmp = litemodule.MyHyperModel()
+        except Exception as e:
+            logger.exception(f'Error: model build (MyHyperModel) in {model_name} failed: {e}')
+            raise SystemExit(1)
+
+        if tmp.batch_size != batch_size:
+            logger.info(f'Overriding model default batch_size of {tmp.batch_size} with batch_size of {batch_size}')
+
+        if tmp.timesteps != 0 and timesteps == 0:
+            timesteps = tmp.timesteps
+            logger.warning(f'Using model default timesteps of {timesteps}')
+
+        model = litemodule.MyHyperModel(timesteps=timesteps, batch_size=batch_size)
+
+    logger.info(f'Loading weights from {ckpt_name}')
+    model.load_state_dict(checkpoint["state_dict"])
+    model.eval()
+    return model
+
+
+def get_frames_from_device(input_name: str | None, length: float, chunk: int = 1024) -> AudioT:
+    p = pyaudio.PyAudio()
+
+    input_devices = get_input_devices(p)
+    if not input_devices:
+        raise ValueError('No input audio devices found')
+
+    if input_name is None:
+        input_name = input_devices[0]
+
+    try:
+        device_index = get_input_device_index_by_name(p, input_name)
+    except ValueError:
+        msg = f'Could not find {input_name}\n'
+        msg += f'Available devices:\n'
+        for input_device in input_devices:
+            msg += f'  {input_device}\n'
+        raise ValueError(msg)
+
+    logger.info(f'Capturing from {p.get_device_info_by_index(device_index).get("name")}')
+    stream = p.open(format=pyaudio.paFloat32,
+                    channels=CHANNEL_COUNT,
+                    rate=SAMPLE_RATE,
+                    input=True,
+                    input_device_index=device_index)
+    stream.start_stream()
+
+    print()
+    print('+---------------------------------+')
+    print('| Press Enter to stop             |')
+    print('+---------------------------------+')
+    print()
+
+    elapsed = 0.0
+    seconds_per_chunk = float(chunk) / float(SAMPLE_RATE)
+    raw_frames = []
+    while elapsed < length or length == -1:
+        raw_frames.append(stream.read(num_frames=chunk, exception_on_overflow=False))
+        elapsed += seconds_per_chunk
+        if select([stdin, ], [], [], 0)[0]:
+            stdin.read(1)
+            length = elapsed
+
+    stream.stop_stream()
+    stream.close()
+    p.terminate()
+    frames = np.frombuffer(b''.join(raw_frames), dtype=np.float32)
+    return frames
+
+
+def get_frames_from_file(input_name: str, length: float) -> AudioT:
+    logger.info(f'Capturing from {input_name}')
+    frames = read_audio(input_name)
+    if length != -1:
+        num_frames = int(length * SAMPLE_RATE)
+        if len(frames) > num_frames:
+            frames = frames[:num_frames]
+    return frames
+
+
+if __name__ == '__main__':
+    main()

sonusai/calc_metric_spenh.py

@@ -978,11 +978,11 @@ def _process_mixture(mixid: int) -> tuple[pd.DataFrame, pd.DataFrame]:
         plot_fname = base_name + '_metric_spenh.pdf'
 
         # Reshape feature to eliminate overlap redundancy for easier to understand spectrogram view
-        # Original size (frames, stride, num_bands), decimates in stride dimension only if step is > 1
-        # Reshape to get frames*decimated_stride, num_bands
+        # Original size (frames, stride, feature_parameters), decimates in stride dimension only if step is > 1
+        # Reshape to get frames*decimated_stride, feature_parameters
         step = int(mixdb.feature_samples / mixdb.feature_step_samples)
         if feature.ndim != 3:
-            raise SonusAIError(f'feature does not have 3 dimensions: frames, stride, num_bands')
+            raise SonusAIError(f'feature does not have 3 dimensions: frames, stride, feature_parameters')
 
         # for feature cn*00n**
         feat_sgram = unstack_complex(feature)

sonusai/data_generator/dataset_from_mixdb.py

@@ -42,7 +42,7 @@ class DatasetFromMixtureDatabase(Sequence):
         self.add1ch = add1ch
         self.shuffle = shuffle
         self.stride = self.mixdb.fg_stride
-        self.num_bands = self.mixdb.fg_num_bands
+        self.feature_parameters = self.mixdb.feature_parameters
         self.num_classes = self.mixdb.num_classes
         self.mixture_frame_segments = None
         self.batch_frame_segments = None

sonusai/data_generator/keras_from_mixdb.py

@@ -61,7 +61,7 @@ class KerasFromMixtureDatabase(Sequence):
         self.add1ch = add1ch
         self.shuffle = shuffle
         self.stride = self.mixdb.fg_stride
-        self.num_bands = self.mixdb.fg_num_bands
+        self.feature_parameters = self.mixdb.feature_parameters
         self.num_classes = self.mixdb.num_classes
         self.mixture_frame_segments: Optional[int] = None
         self.batch_frame_segments: Optional[int] = None

sonusai/genft.py

@@ -165,7 +165,8 @@ def main() -> None:
     logger.info(f'Wrote {len(mixids)} mixtures to {location}')
     logger.info('')
     logger.info(f'Duration: {seconds_to_hms(seconds=duration)}')
-    logger.info(f'feature:  {human_readable_size(total_feature_frames * mixdb.fg_stride * mixdb.fg_num_bands * 4, 1)}')
+    logger.info(
+        f'feature:  {human_readable_size(total_feature_frames * mixdb.fg_stride * mixdb.feature_parameters * 4, 1)}')
     logger.info(f'truth_f:  {human_readable_size(total_feature_frames * mixdb.num_classes * 4, 1)}')
     if compute_segsnr:
         logger.info(f'segsnr:   {human_readable_size(total_transform_frames * 4, 1)}')

sonusai/genmixdb.py

@@ -337,12 +337,12 @@ def genmixdb(location: str,
         log_duration_and_sizes(total_duration=total_duration,
                                num_classes=mixdb.num_classes,
                                feature_step_samples=mixdb.feature_step_samples,
-                               num_bands=mixdb.fg_num_bands,
+                               feature_parameters=mixdb.feature_parameters,
                                stride=mixdb.fg_stride,
                                desc='Estimated')
         logger.info(f'Feature shape:        '
-                    f'{mixdb.fg_stride} x {mixdb.fg_num_bands} '
-                    f'({mixdb.fg_stride * mixdb.fg_num_bands} total params)')
+                    f'{mixdb.fg_stride} x {mixdb.feature_parameters} '
+                    f'({mixdb.fg_stride * mixdb.feature_parameters} total params)')
         logger.info(f'Feature samples:      {mixdb.feature_samples} samples ({mixdb.feature_ms} ms)')
         logger.info(f'Feature step samples: {mixdb.feature_step_samples} samples ({mixdb.feature_step_ms} ms)')
         logger.info('')
@@ -371,7 +371,7 @@ def genmixdb(location: str,
         log_duration_and_sizes(total_duration=total_duration,
                                num_classes=mixdb.num_classes,
                                feature_step_samples=mixdb.feature_step_samples,
-                               num_bands=mixdb.fg_num_bands,
+                               feature_parameters=mixdb.feature_parameters,
                                stride=mixdb.fg_stride,
                                desc='Actual')
         logger.info('')

sonusai/keras_predict.py

@@ -180,7 +180,7 @@ def main() -> None:
     for file in input_name:
         # Convert WAV to feature data
         audio = read_audio(file)
-        feature = get_feature_from_audio(audio=audio, feature=hypermodel.feature)
+        feature = get_feature_from_audio(audio=audio, feature_mode=hypermodel.feature)
 
         feature, predict = _pad_and_predict(hypermodel=hypermodel,
                                             built_model=built_model,

sonusai/lsdb.py

@@ -48,8 +48,8 @@ def lsdb(mixdb: MixtureDatabase,
     logger.info(f'{"Targets":{desc_len}} {mixdb.num_target_files}')
     logger.info(f'{"Noises":{desc_len}} {mixdb.num_noise_files}')
     logger.info(f'{"Feature":{desc_len}} {mixdb.feature}')
-    logger.info(f'{"Feature shape":{desc_len}} {mixdb.fg_stride} x {mixdb.fg_num_bands} '
-                f'({mixdb.fg_stride * mixdb.fg_num_bands} total params)')
+    logger.info(f'{"Feature shape":{desc_len}} {mixdb.fg_stride} x {mixdb.feature_parameters} '
+                f'({mixdb.fg_stride * mixdb.feature_parameters} total params)')
     logger.info(f'{"Feature samples":{desc_len}} {mixdb.feature_samples} samples ({mixdb.feature_ms} ms)')
     logger.info(f'{"Feature step samples":{desc_len}} {mixdb.feature_step_samples} samples '
                 f'({mixdb.feature_step_ms} ms)')

sonusai/main.py

@@ -3,9 +3,9 @@
 usage: sonusai [--version] [--help] <command> [<args>...]
 
 The sonusai commands are:
+   audiofe                      Audio front end
    calc_metric_spenh            Run speech enhancement and analysis
    doc                          Documentation
-   evaluate                     Evaluate model performance
    genft                        Generate feature and truth data
    genmix                       Generate mixture and truth data
    genmixdb                     Generate a mixture database
@@ -39,9 +39,9 @@ def main() -> None:
     from sonusai.utils import trim_docstring
 
     commands = (
+        'audiofe',
         'calc_metric_spenh',
         'doc',
-        'evaluate',
         'genft',
         'genmix',
         'genmixdb',

sonusai/mixture/__init__.py

@@ -1,6 +1,4 @@
 # SonusAI mixture utilities
-from .audio import calculate_audio_from_transform
-from .audio import calculate_transform_from_audio
 from .audio import get_duration
 from .audio import get_next_noise
 from .audio import get_num_samples
@@ -83,6 +81,7 @@ from .datatypes import TruthFunctionConfig
 from .datatypes import TruthSetting
 from .datatypes import TruthSettings
 from .datatypes import UniversalSNR
+from .feature import get_audio_from_feature
 from .feature import get_feature_from_audio
 from .generation import generate_mixtures
 from .generation import get_all_snrs_from_config
@@ -102,8 +101,10 @@ from .helpers import augmented_noise_samples
 from .helpers import augmented_target_samples
 from .helpers import check_audio_files_exist
 from .helpers import forward_transform
+from .helpers import get_audio_from_transform
 from .helpers import get_ft
 from .helpers import get_segsnr
+from .helpers import get_transform_from_audio
 from .helpers import get_truth_t
 from .helpers import inverse_transform
 from .helpers import mixture_metadata

sonusai/mixture/audio.py

@@ -1,11 +1,6 @@
 from functools import lru_cache
 
-from pyaaware import ForwardTransform
-from pyaaware import InverseTransform
-
-from sonusai.mixture.datatypes import AudioF
 from sonusai.mixture.datatypes import AudioT
-from sonusai.mixture.datatypes import EnergyT
 from sonusai.mixture.datatypes import ImpulseResponseData
 
 
@@ -22,35 +17,6 @@ def get_next_noise(audio: AudioT, offset: int, length: int) -> AudioT:
     return np.take(audio, range(offset, offset + length), mode='wrap')
 
 
-def calculate_transform_from_audio(audio: AudioT,
-                                   transform: ForwardTransform) -> tuple[AudioF, EnergyT]:
-    """Apply forward transform to input audio data to generate transform data
-
-    :param audio: Time domain data [samples]
-    :param transform: ForwardTransform object
-    :return: Frequency domain data [frames, bins], Energy [frames]
-    """
-    f, e = transform.execute_all(audio)
-    return f.transpose(), e
-
-
-def calculate_audio_from_transform(data: AudioF,
-                                   transform: InverseTransform,
-                                   trim: bool = True) -> tuple[AudioT, EnergyT]:
-    """Apply inverse transform to input transform data to generate audio data
-
-    :param data: Frequency domain data [frames, bins]
-    :param transform: InverseTransform object
-    :param trim: Removes starting samples so output waveform will be time-aligned with input waveform to the transform
-    :return: Time domain data [samples], Energy [frames]
-    """
-    t, e = transform.execute_all(data.transpose())
-    if trim:
-        t = t[transform.N - transform.R:]
-
-    return t, e
-
-
 def get_duration(audio: AudioT) -> float:
     """Get duration of audio in seconds
 

sonusai/mixture/datatypes.py

@@ -304,7 +304,7 @@ class FeatureGeneratorInfo:
     decimation: int
     stride: int
     step: int
-    num_bands: int
+    feature_parameters: int
     ft_config: TransformConfig
     eft_config: TransformConfig
     it_config: TransformConfig

sonusai/mixture/feature.py

@@ -1,51 +1,105 @@
+from typing import Optional
+
+from sonusai.mixture.datatypes import AudioF
 from sonusai.mixture.datatypes import AudioT
 from sonusai.mixture.datatypes import Feature
 
 
-def get_feature_from_audio(audio: AudioT, feature: str) -> Feature:
-    from dataclasses import asdict
+def get_feature_from_audio(audio: AudioT,
+                           feature_mode: str,
+                           num_classes: Optional[int] = 1,
+                           truth_mutex: Optional[bool] = False) -> Feature:
+    """Apply forward transform and generate feature data from audio data
 
+    :param audio: Time domain audio data [samples]
+    :param feature_mode: Feature mode
+    :param num_classes: Number of classes
+    :param truth_mutex: Whether to calculate 'other' label
+    :return: Feature data [frames, strides, feature_parameters]
+    """
     import numpy as np
     from pyaaware import FeatureGenerator
 
     from .augmentation import pad_audio_to_frame
-    from .datatypes import FeatureGeneratorConfig
     from .datatypes import TransformConfig
     from .helpers import forward_transform
-    from .truth import truth_reduction
 
-    num_classes = 1
-    truth_mutex = False
-    truth_reduction_function = 'max'
+    fg = FeatureGenerator(feature_mode=feature_mode,
+                          num_classes=num_classes,
+                          truth_mutex=truth_mutex)
 
-    fg_config = FeatureGeneratorConfig(feature_mode=feature,
-                                       num_classes=num_classes,
-                                       truth_mutex=truth_mutex)
-    fg = FeatureGenerator(**asdict(fg_config))
     feature_step_samples = fg.ftransform_R * fg.decimation * fg.step
-
     audio = pad_audio_to_frame(audio, feature_step_samples)
-    samples = len(audio)
-    audio_f = forward_transform(audio, TransformConfig(N=fg.ftransform_N,
+
+    audio_f = forward_transform(audio=audio,
+                                config=TransformConfig(N=fg.ftransform_N,
                                                        R=fg.ftransform_R,
                                                        bin_start=fg.bin_start,
                                                        bin_end=fg.bin_end,
                                                        ttype=fg.ftransform_ttype))
 
+    samples = len(audio)
     transform_frames = samples // fg.ftransform_R
     feature_frames = samples // feature_step_samples
 
-    truth_t = np.empty((samples, num_classes), dtype=np.float32)
-
-    data = np.empty((feature_frames, fg.stride, fg.num_bands), dtype=np.float32)
+    feature = np.empty((feature_frames, fg.stride, fg.feature_parameters), dtype=np.float32)
 
     feature_frame = 0
     for transform_frame in range(transform_frames):
-        indices = slice(transform_frame * fg.ftransform_R, (transform_frame + 1) * fg.ftransform_R)
-        fg.execute(audio_f[transform_frame], truth_reduction(truth_t[indices], truth_reduction_function))
+        fg.execute(audio_f[transform_frame])
 
         if fg.eof():
-            data[feature_frame] = fg.feature()
+            feature[feature_frame] = fg.feature()
             feature_frame += 1
 
-    return data
+    return feature
+
+
+def get_audio_from_feature(feature: Feature,
+                           feature_mode: str,
+                           num_classes: Optional[int] = 1,
+                           truth_mutex: Optional[bool] = False,
+                           trim: Optional[bool] = True) -> AudioT:
+    """Apply inverse transform to feature data to generate audio data
+
+    :param feature: Feature data [frames, strides, feature_parameters]
+    :param feature_mode: Feature mode
+    :param num_classes: Number of classes
+    :param truth_mutex: Whether to calculate 'other' label
+    :param trim: Whether to trim the audio data
+    :return: Audio data [samples]
+    """
+    import numpy as np
+
+    from pyaaware import FeatureGenerator
+
+    from .datatypes import TransformConfig
+    from .helpers import inverse_transform
+    from sonusai.utils.stacked_complex import unstack_complex
+
+    fg = FeatureGenerator(feature_mode=feature_mode,
+                          num_classes=num_classes,
+                          truth_mutex=truth_mutex)
+
+    feature_complex = unstack_complex(feature)
+    if feature_mode[0:1] == 'h':
+        feature_complex = _power_uncompress(feature_complex)
+    return np.squeeze(inverse_transform(transform=feature_complex,
+                                        config=TransformConfig(N=fg.itransform_N,
+                                                               R=fg.itransform_R,
+                                                               bin_start=fg.bin_start,
+                                                               bin_end=fg.bin_end,
+                                                               ttype=fg.itransform_ttype),
+                                        trim=trim))
+
+
+def _power_uncompress(feature: AudioF) -> AudioF:
+    import numpy as np
+
+    mag = np.abs(feature)
+    phase = np.angle(feature)
+    mag = mag ** (1. / 0.3)
+    real_uncompress = mag * np.cos(phase)
+    imag_uncompress = mag * np.sin(phase)
+
+    return real_uncompress + 1j * imag_uncompress

sonusai/mixture/generation.py

@@ -433,7 +433,7 @@ def _initialize_targets_audio(mixdb: MixtureDatabase, mixture: Mixture) -> tuple
         # target_gain is used to back out the gain augmentation in order to return the target audio
         # to its normalized level when calculating truth (if needed).
         if target.augmentation.gain is not None:
-            target.gain = round(10 ** (target.augmentation.gain / 20), ndigits=7)
+            target.gain = round(10 ** (target.augmentation.gain / 20), ndigits=5)
         else:
             target.gain = 1
 
@@ -507,8 +507,8 @@ def _initialize_mixture_gains(mixdb: MixtureDatabase,
         mixture.target_snr_gain *= gain_adjustment
         mixture.noise_snr_gain *= gain_adjustment
 
-    mixture.target_snr_gain = round(mixture.target_snr_gain, ndigits=7)
-    mixture.noise_snr_gain = round(mixture.noise_snr_gain, ndigits=7)
+    mixture.target_snr_gain = round(mixture.target_snr_gain, ndigits=5)
+    mixture.noise_snr_gain = round(mixture.noise_snr_gain, ndigits=5)
     return mixture