Skip to content

Latest commit

 

History

History
99 lines (82 loc) · 3.21 KB

ASRSupported.md

File metadata and controls

99 lines (82 loc) · 3.21 KB

Supported Model architecture

Encoder

arch name supported
ConformerEncoder Yes
ContextualBlockConformerEncoder Yes
ContextualBlockTransformerEncoderx Yes
FairseqHubertEncoder No
RNNEncoder Yes
TransformerEncoder Yes
VGGRNNEncoderx Yes
FairSeqWav2Vec2Encoder No
FairseqHubertEncoder No
FairseqHubertPretrainEncoder No
LongformerEncoder No
BranchformerEncoder Yes
E-BranchformerEncoder Yes

Decoder

arch name supported
RNNDecoder Yes
TransformerDecoder Yes
LightweightConvolutionTransformerDecoder Yes
LightweightConvolution2DTransformerDecoder Yes
DynamicConvolutionTransformerDecoder No
DynamicConvolution2DTransformerDecoder No
TransducerDecoder Yes
MLMDecoder No

Language Model

arch name supported
SequentialRNNLM Yes
TransformerLM Yes

frontend

arch name supported
DefaultFrontend Yes
SlidingWindow No
HuBERT Yes
Wav2Vec No
FusedFrontends No
No Frontend Yes

normalize

arch name supported
GlobalMVN Yes
UtteranceMVN Yes

PositionalEmbedding

arch name supported
PositionalEncoding Yes
ScaledPositionalEncoding Yes
LegacyRelPositionalEncoding Yes
RelPositionalEncoding Yes
StreamPositionalEncoding Yes

Attention

arch name supported
NoAtt Yes
AttDot Yes
AttAdd Yes
AttLoc Yes
AttCov Yes
AttLoc2D No
AttLocRec No
AttCovLoc Yes
AttMultiHeadDot No
AttMultiHeadAdd No
AttMultiHeadLoc No
AttMultiHeadMultiResLoc No
AttForward No
AttForwardTA No

pre encoder The following shows the encoder module that support preencoder export and inference.

arch name supported
RNN Yes
Transformer Yes
Conformer Yes
ContextualBlock No

post encoder not supported.