Models#

This section gives a brief overview of the models that NeMo’s ASR collection currently supports.

Each of these models can be used with the example ASR scripts (in the <NeMo_git_root>/examples/asr directory) by specifying the model architecture in the config file used. Examples of config files for each model can be found in the <NeMo_git_root>/examples/asr/conf directory.

For more information about the config files and how they should be structured, refer to the NeMo ASR Configuration Files section.

Pretrained checkpoints for all of these models, as well as instructions on how to load them, can be found in the Checkpoints section. You can use the available checkpoints for immediate inference, or fine-tune them on your own datasets. The checkpoints section also contains benchmark results for the available ASR models.

Jasper#

Jasper (“Just Another Speech Recognizer”) [ASR-MODELS6] is a deep time delay neural network (TDNN) comprising of blocks of 1D-convolutional layers. The Jasper family of models are denoted as Jasper_[BxR] where B is the number of blocks and R is the number of convolutional sub-blocks within a block. Each sub-block contains a 1-D convolution, batch normalization, ReLU, and dropout:

jasper model

Jasper models can be instantiated using the EncDecCTCModel class.

QuartzNet#

QuartzNet [ASR-MODELS5] is a version of Jasper [ASR-MODELS6] model with separable convolutions and larger filters. It can achieve performance similar to Jasper but with an order of magnitude fewer parameters. Similarly to Jasper, the QuartzNet family of models are denoted as QuartzNet_[BxR] where B is the number of blocks and R is the number of convolutional sub-blocks within a block. Each sub-block contains a 1-D separable convolution, batch normalization, ReLU, and dropout:

quartznet model

QuartzNet models can be instantiated using the EncDecCTCModel class.

Citrinet#

Citrinet is a version of QuartzNet [ASR-MODELS5] that extends ContextNet [ASR-MODELS2], utilizing subword encoding (via Word Piece tokenization) and Squeeze-and-Excitation mechanism [ASR-MODELS4] to obtain highly accurate audio transcripts while utilizing a non-autoregressive CTC based decoding scheme for efficient inference.

citrinet model

Citrinet models can be instantiated using the EncDecCTCModelBPE class.

ContextNet#

ContextNet is a model uses Transducer/RNNT loss/decoder and is introduced in [ASR-MODELS2]. It uses Squeeze-and-Excitation mechanism [ASR-MODELS4] to model larger context. Unlike Citrinet, it has an autoregressive decoding scheme.

ContextNet models can be instantiated using the EncDecRNNTBPEModel class for a model with sub-word encoding and EncDecRNNTModel for char-based encoding.

You may find the example config files of ContextNet model with character-based encoding at <NeMo_git_root>/examples/asr/conf/contextnet_rnnt/contextnet_rnnt_char.yaml and with sub-word encoding at <NeMo_git_root>/examples/asr/conf/contextnet_rnnt/contextnet_rnnt.yaml.

Conformer-CTC#

Conformer-CTC is a CTC-based variant of the Conformer model introduced in [ASR-MODELS1]. Conformer-CTC has a similar encoder as the original Conformer but uses CTC loss and decoding instead of RNNT/Transducer loss, which makes it a non-autoregressive model. We also drop the LSTM decoder and instead use a linear decoder on the top of the encoder. This model uses the combination of self-attention and convolution modules to achieve the best of the two approaches, the self-attention layers can learn the global interaction while the convolutions efficiently capture the local correlations. The self-attention modules support both regular self-attention with absolute positional encoding, and also Transformer-XL’s self-attention with relative positional encodings.

Here is the overall architecture of the encoder of Conformer-CTC:

Conformer-CTC Model

This model supports both the sub-word level and character level encodings. You can find more details on the config files for the Conformer-CTC models at Conformer-CTC <./configs.html#conformer-ctc>. The variant with sub-word encoding is a BPE-based model which can be instantiated using the EncDecCTCModelBPE class, while the character-based variant is based on EncDecCTCModel.

You may find the example config files of Conformer-CTC model with character-based encoding at <NeMo_git_root>/examples/asr/conf/conformer/conformer_ctc_char.yaml and with sub-word encoding at <NeMo_git_root>/examples/asr/conf/conformer/conformer_ctc_bpe.yaml.

Conformer-Transducer#

Conformer-Transducer is the Conformer model introduced in [ASR-MODELS1] and uses RNNT/Transducer loss/decoder. It has the same encoder as Conformer-CTC but utilizes RNNT/Transducer loss/decoder which makes it an autoregressive model.

Most of the config file for Conformer-Transducer models are similar to Conformer-CTC except the sections related to the decoder and loss: decoder, loss, joint, decoding. You may take a look at our tutorials page <../starthere/tutorials.html> on Transducer models to become familiar with their configs: Introduction to Transducers <https://colab.research.google.com/github/NVIDIA/NeMo/blob/stable/tutorials/asr/Intro_to_Transducers.ipynb> and ASR with Transducers <https://colab.research.google.com/github/NVIDIA/NeMo/blob/stable/tutorials/asr/ASR_with_Transducers.ipynb> You can find more details on the config files for the Conformer-Transducer models at Conformer-CTC <./configs.html#conformer-ctc>.

This model supports both the sub-word level and character level encodings. The variant with sub-word encoding is a BPE-based model which can be instantiated using the EncDecRNNTBPEModel class, while the character-based variant is based on EncDecRNNTModel.

You may find the example config files of Conformer-Transducer model with character-based encoding at <NeMo_git_root>/examples/asr/conf/conformer/conformer_transducer_char.yaml and with sub-word encoding at <NeMo_git_root>/examples/asr/conf/conformer/conformer_transducer_bpe.yaml.

LSTM-Transducer#

LSTM-Transducer is a model which uses RNNs (eg. LSTM) in the encoder. The architecture of this model is followed from suggestions in [ASR-MODELS3]. It uses RNNT/Transducer loss/decoder. The encoder consists of RNN layers (LSTM as default) with lower projection size to increase the efficiency. Layer norm is added between the layers to stabilize the training. It can be trained/used in unidirectional or bidirectional mode. The unidirectional mode is fully causal and can be used easily for simple and efficient frame-wise streaming. However the accuracy of this model is generally lower than other models like Conformer and Citrinet.

This model supports both the sub-word level and character level encodings. You may find the example config file of RNNT model with wordpiece encoding at <NeMo_git_root>/examples/asr/conf/lstm/lstm_transducer_bpe.yaml. You can find more details on the config files for the RNNT models at LSTM-Transducer <./configs.html#lstm-transducer>.

LSTM-CTC#

LSTM-CTC model is a CTC-variant of the LSTM-Transducer model which uses CTC loss/decoding instead of Transducer. You may find the example config file of LSTM-CTC model with wordpiece encoding at <NeMo_git_root>/examples/asr/conf/lstm/lstm_ctc_bpe.yaml.

References#

ASR-MODELS1(1,2)

Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and others. Conformer: convolution-augmented transformer for speech recognition. arXiv preprint arXiv:2005.08100, 2020.

ASR-MODELS2(1,2)

Wei Han, Zhengdong Zhang, Yu Zhang, Jiahui Yu, Chung-Cheng Chiu, James Qin, Anmol Gulati, Ruoming Pang, and Yonghui Wu. Contextnet: improving convolutional neural networks for automatic speech recognition with global context. arXiv:2005.03191, 2020.

ASR-MODELS3

Yanzhang He, Tara N Sainath, Rohit Prabhavalkar, Ian McGraw, Raziel Alvarez, Ding Zhao, David Rybach, Anjuli Kannan, Yonghui Wu, Ruoming Pang, and others. Streaming end-to-end speech recognition for mobile devices. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 6381–6385. IEEE, 2019.

ASR-MODELS4(1,2)

Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In ICVPR. 2018.

ASR-MODELS5(1,2)

Samuel Kriman, Stanislav Beliaev, Boris Ginsburg, Jocelyn Huang, Oleksii Kuchaiev, Vitaly Lavrukhin, Ryan Leary, Jason Li, and Yang Zhang. Quartznet: Deep automatic speech recognition with 1d time-channel separable convolutions. arXiv preprint arXiv:1910.10261, 2019.

ASR-MODELS6(1,2)

Jason Li, Vitaly Lavrukhin, Boris Ginsburg, Ryan Leary, Oleksii Kuchaiev, Jonathan M Cohen, Huyen Nguyen, and Ravi Teja Gadde. Jasper: an end-to-end convolutional neural acoustic model. arXiv preprint arXiv:1904.03288, 2019.