class EncDecClassificationConfig(model_cfg.ModelConfig):
# Model global arguments
sample_rate: int = 16000
repeat: int = 1
dropout: float = 0.0
separable: bool = True
kernel_size_factor: float = 1.0
labels: List[str] = MISSING
timesteps: int = MISSING
# Dataset configs
train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig(
manifest_filepath=None, shuffle=True, trim_silence=False)
validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig(
manifest_filepath=None, shuffle=False)
test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig(
manifest_filepath=None, shuffle=False)
# Optimizer / Scheduler config
optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(
sched=model_cfg.SchedConfig())
# Model component configs
preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(
)
spec_augment: Optional[
SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig()
crop_or_pad_augment: Optional[
CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig(
audio_length=timesteps)
encoder: ConvASREncoderConfig = ConvASREncoderConfig()
decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig(
)
class JasperModelConfig(ctc_cfg.EncDecCTCConfig):
# Model global arguments
sample_rate: int = 16000
repeat: int = 1
dropout: float = 0.0
separable: bool = False
labels: List[str] = MISSING
# Dataset configs
train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(
manifest_filepath=None, shuffle=True, trim_silence=True)
validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(
manifest_filepath=None, shuffle=False)
test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(
manifest_filepath=None, shuffle=False)
# Optimizer / Scheduler config
optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(
sched=model_cfg.SchedConfig())
# Model general component configs
preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig(
)
spec_augment: Optional[
SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig()
encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu")
decoder: ConvASRDecoderConfig = ConvASRDecoderConfig()
class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig):
# Model global arguments
sample_rate: int = 16000
repeat: int = 1
dropout: float = 0.0
separable: bool = True
kernel_size_factor: float = 1.0
timesteps: int = 128
labels: List[str] = MISSING
# Dataset configs
train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig(
manifest_filepath=None, shuffle=True, trim_silence=False)
validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig(
manifest_filepath=None, shuffle=False)
test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig(
manifest_filepath=None, shuffle=False)
# Optimizer / Scheduler config
optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(
sched=model_cfg.SchedConfig())
# Model general component configs
preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(
window_size=0.025)
spec_augment: Optional[
SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig(
freq_masks=2,
time_masks=2,
freq_width=15,
time_width=25,
rect_masks=5,
rect_time=25,
rect_freq=15)
crop_or_pad_augment: Optional[
CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig(
audio_length=128)
encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu")
decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig(
)
