from malaya_speech.supervised import unet
from herpetologist import check_type
from malaya_speech.utils import describe_availability
import logging
logger = logging.getLogger(__name__)
# https://github.com/sigsep/sigsep-mus-eval/blob/master/museval/__init__.py#L364
# Only calculate SDR, ISR, SAR on voice sample
_masking_availability = {
'unet': {
'Size (MB)': 78.9,
'Quantized Size (MB)': 20,
'SUM MAE': 0.85896,
'MAE_SPEAKER': 0.46849,
'MAE_NOISE': 0.39046,
'SDR': 12.12805,
'ISR': 14.67067,
'SAR': 15.019682,
},
'resnet-unet': {
'Size (MB)': 91.4,
'Quantized Size (MB)': 23,
'SUM MAE': 0.8153995,
'MAE_SPEAKER': 0.447958,
'MAE_NOISE': 0.367441,
'SDR': 12.349259,
'ISR': 14.85418,
'SAR': 15.21751,
},
}
_sampling_availability = {
'unet': {
'Size (MB)': 40.7,
'Quantized Size (MB)': 10.3,
'SDR': 9.877178,
'ISR': 15.916217,
'SAR': 13.709130,
},
'resnet-unet': {
'Size (MB)': 36.4,
'Quantized Size (MB)': 9.29,
'SDR': 9.43617,
'ISR': 16.86103,
'SAR': 12.32157,
},
'resnext-unet': {
'Size (MB)': 36.1,
'Quantized Size (MB)': 9.26,
'SDR': 9.685578,
'ISR': 16.42137,
'SAR': 12.45115,
},
}
[docs]def available_deep_masking():
"""
List available Speech Enhancement STFT masking deep learning model.
"""
logger.info('Only calculate SDR, ISR, SAR on voice sample. Higher is better.')
return describe_availability(_masking_availability)
[docs]def available_deep_enhance():
"""
List available Speech Enhancement UNET Waveform sampling deep learning model.
"""
logger.info('Only calculate SDR, ISR, SAR on voice sample. Higher is better.')
return describe_availability(_sampling_availability)
[docs]@check_type
def deep_masking(model: str = 'resnet-unet', quantized: bool = False, **kwargs):
"""
Load Speech Enhancement STFT UNET masking deep learning model.
Parameters
----------
model : str, optional (default='resnet-unet')
Check available models at `malaya_speech.speech_enhancement.available_deep_masking()`.
quantized : bool, optional (default=False)
if True, will load 8-bit quantized model.
Quantized model not necessary faster, totally depends on the machine.
Returns
-------
result : malaya_speech.model.unet.UNETSTFT class
"""
model = model.lower()
if model not in _masking_availability:
raise ValueError(
'model not supported, please check supported models from `malaya_speech.speech_enhancement.available_deep_masking()`.'
)
return unet.load_stft(
model=model,
module='speech-enhancement-mask',
instruments=['voice', 'noise'],
quantized=quantized,
**kwargs
)
[docs]@check_type
def deep_enhance(model: str = 'unet', quantized: bool = False, **kwargs):
"""
Load Speech Enhancement UNET Waveform sampling deep learning model.
Parameters
----------
model : str, optional (default='unet')
Check available models at `malaya_speech.speech_enhancement.available_deep_enhance()`.
quantized : bool, optional (default=False)
if True, will load 8-bit quantized model.
Quantized model not necessary faster, totally depends on the machine.
Returns
-------
result : malaya_speech.model.unet.UNET1D class
"""
model = model.lower()
if model not in _sampling_availability:
raise ValueError(
'model not supported, please check supported models from `malaya_speech.speech_enhancement.available_deep_enhance()`.'
)
return unet.load_1d(
model=model,
module='speech-enhancement',
quantized=quantized,
**kwargs
)
