from malaya_speech.supervised import classification
from malaya_speech.utils import describe_availability
from herpetologist import check_type
import logging
import warnings
logger = logging.getLogger(__name__)
# EER calculation, https://github.com/huseinzol05/malaya-speech/tree/master/pretrained-model/speaker-embedding/calculate-EER
# EER tested on VoxCeleb2 test set.
_availability = {
'deep-speaker': {
'Size (MB)': 96.7,
'Quantized Size (MB)': 24.4,
'Embedding Size': 512,
'EER': 0.2187,
},
'vggvox-v1': {
'Size (MB)': 70.8,
'Quantized Size (MB)': 17.7,
'Embedding Size': 1024,
'EER': 0.13944,
},
'vggvox-v2': {
'Size (MB)': 43.2,
'Quantized Size (MB)': 7.92,
'Embedding Size': 512,
'EER': 0.0446,
},
'conformer-base': {
'Size (MB)': 99.4,
'Quantized Size (MB)': 27.2,
'Embedding Size': 512,
'EER': 0.06938,
},
'conformer-tiny': {
'Size (MB)': 20.3,
'Quantized Size (MB)': 6.21,
'Embedding Size': 512,
'EER': 0.08687,
},
}
trillsson_accuracy = {
'trillsson-1': {
'url': 'https://tfhub.dev/google/nonsemantic-speech-benchmark/trillsson1/1',
'EER': 0.3804599,
},
'trillsson-2': {
'url': 'https://tfhub.dev/google/nonsemantic-speech-benchmark/trillsson2/1',
'EER': 0.3898799,
}
}
dvector_accuracy = {
'original from': 'https://github.com/yistLin/dvector',
'Size (MB)': 5.45,
'Embedding Size': 256,
'EER': 0.1356490598298,
}
_nemo_availability = {
'huseinzol05/nemo-ecapa-tdnn': {
'Size (MB)': 96.8,
'Embedding Size': 192,
'EER': 0.0249200000000007,
'original from': 'https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/ecapa_tdnn',
},
'huseinzol05/nemo-speakernet': {
'Size (MB)': 23.6,
'Embedding Size': 192,
'EER': 0.0427898305,
'original from': 'https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/speakerverification_speakernet',
},
'huseinzol05/nemo-titanet_large': {
'Size (MB)': 101.6,
'Embedding Size': 192,
'EER': 0.02277999999996,
'original from': 'https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/titanet_large',
}
}
_huggingface_availability = {
'microsoft/wavlm-base-sv': {
'Size (MB)': 405,
'Embedding Size': 512,
'EER': 0.07827375115,
},
'microsoft/wavlm-base-plus-sv': {
'Size (MB)': 405,
'Embedding Size': 512,
'EER': 0.06688427572,
},
'microsoft/unispeech-sat-large-sv': {
'Size (MB)': 1290,
'Embedding Size': 512,
'EER': 0.2032767553,
},
'microsoft/unispeech-sat-base-sv': {
'Size (MB)': 404,
'Embedding Size': 512,
'EER': 0.0782815656,
},
'microsoft/unispeech-sat-base-plus-sv': {
'Size (MB)': 404,
'Embedding Size': 512,
'EER': 0.0761281698,
},
}
def _describe():
logger.info('tested on VoxCeleb2 test set. Lower EER is better.')
logger.info('download the test set at https://github.com/huseinzol05/malaya-speech/tree/master/data/voxceleb')
[docs]def available_model():
"""
List available speaker vector deep models using Tensorflow.
"""
_describe()
return describe_availability(_availability)
[docs]def available_nemo():
"""
List available Nvidia Nemo Speaker vector models.
"""
_describe()
return describe_availability(_nemo_availability)
[docs]def available_huggingface():
"""
List available HuggingFace Speaker vector models.
"""
_describe()
return describe_availability(_huggingface_availability)
[docs]@check_type
def deep_model(model: str = 'vggvox-v2', quantized: bool = False, **kwargs):
"""
Load Speaker2Vec model.
Parameters
----------
model : str, optional (default='speakernet')
Check available models at `malaya_speech.speaker_vector.available_model()`.
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.supervised.classification.load function
"""
model = model.lower()
if model not in _availability:
raise ValueError(
'model not supported, please check supported models from `malaya_speech.speaker_vector.available_model()`.'
)
return classification.load(
model=model,
module='speaker-vector',
extra={},
label=None,
quantized=quantized,
**kwargs
)
[docs]@check_type
def nemo(
model: str = 'huseinzol05/nemo-ecapa-tdnn',
**kwargs,
):
"""
Load Nemo Speaker verification model.
Parameters
----------
model : str, optional (default='huseinzol05/nemo-ecapa-tdnn')
Check available models at `malaya_speech.speaker_vector.available_nemo()`.
Returns
-------
result : malaya_speech.torch_model.nemo.SpeakerVector class
"""
if model not in _nemo_availability:
raise ValueError(
'model not supported, please check supported models from `malaya_speech.speaker_vector.available_nemo()`.'
)
return classification.nemo_speaker_vector(
model=model,
**kwargs
)
[docs]@check_type
def huggingface(
model: str = 'microsoft/wavlm-base-plus-sv',
force_check: bool = True,
**kwargs,
):
"""
Load Finetuned models from HuggingFace.
Parameters
----------
model : str, optional (default='microsoft/wavlm-base-plus-sv')
Check available models at `malaya_speech.speaker_vector.available_huggingface()`.
force_check: bool, optional (default=True)
Force check model one of malaya model.
Set to False if you have your own huggingface model.
Returns
-------
result : malaya_speech.torch_model.huggingface.XVector class
"""
if model not in _huggingface_availability and force_check:
raise ValueError(
'model not supported, please check supported models from `malaya_speech.speaker_vector.available_huggingface()`.'
)
return classification.huggingface_xvector(
model=model,
**kwargs
)
