Cascaded encoders for simplified streaming and non-streaming ASR

1. An automated speech recognition (ASR) model comprising:

a first encoder configured to:

receive, as input, a sequence of acoustic frames; and

generate, at each of a plurality of output steps, a first higher order feature representation for a corresponding acoustic frame in the sequence of acoustic frames;

a second encoder configured to:

receive, as input, the first higher order feature representation generated by the first encoder at each of the plurality of output steps; and

generate, at each of the plurality of output steps, a second higher order feature representation for a corresponding first higher order feature frame; and

a single decoder configured to:

when the ASR model is operating in a non-streaming mode:

receive, as input, the second higher order feature representation generated by the second encoder at each of the plurality of output steps; and

generate, at each of the plurality of output steps, a first probability distribution over possible speech recognition hypotheses; and

when the ASR model is operating in a streaming mode:

receive, as input, the first higher order feature representation generated by the first encoder at each of the plurality of output steps; and

generate, at each of the plurality of output steps, a second probability distribution over possible speech recognition hypotheses;

wherein a training process trains the model to learn how to operate in the non-streaming mode and the streaming mode by:

receiving a set of training utterances; and

jointly training, in a single stage, the ASR model using the set of training utterances by:

for each training utterance of the set of training utterances:

sampling the training utterance of the set of training utterances;

stochastically selecting one of the non-streaming mode of the ASR model and the streaming mode of the ASR model;

training the selected one of the non-streaming mode of the ASR model and the streaming mode of the ASR model; and

calculating a loss for the training utterance only once;

generating a first loss for the first encoder;

generating a second loss for the second encoder; and

minimizing a weighted sum of the first loss and the second loss.

2. The ASR model of claim 1 , wherein the second encoder generates the second higher order feature representation without receiving any of the acoustic frames as input.

3. The ASR model of claim 1 , wherein the decoder is further configured to generate partial speech recognition results based on the second probability distribution over possible speech recognition hypotheses.

4. The ASR model of claim 1 , wherein the single decoder comprises:

a prediction network configured to:

receive, as input, a sequence of non-blank symbols output by a final Softmax layer; and

generate, at each of the plurality of output steps, a dense representation; and

a joint network configured to:

receive, as input, the dense representation generated by the prediction network at each of the plurality of output steps and one of:

when the ASR model is operating in the streaming mode, the first higher order feature representation generated by the first encoder at each of the plurality of output steps; or

when the ASR model is operating in the non-streaming mode, the second higher order feature representation generated by the second encoder at each of the plurality of output steps; and

generate, at each of the plurality of output steps, one of:

when the ASR model is operating in the streaming mode, the second probability distribution over possible speech recognition hypotheses; or

when the ASR model is operating in the non-streaming mode, the first probability distribution over possible speech recognition hypotheses.

5. The ASR model of claim 4 , wherein the prediction network comprises a long short-term memory (LSTM)-based prediction network.

6. The ASR model of claim 4 , wherein the prediction network comprises a V2 embedding look-up table.

7. The ASR model of claim 1 , wherein the first encoder comprises a causal encoder comprising one of:

a plurality of unidirectional long short-term memory (LSTM) layers;

a plurality of conformer layers; or

a plurality of transformer layers.

8. The ASR model of claim 1 , wherein the second encoder comprises a non-causal encoder comprising one of:

one or more bi-directional long short-term memory (LSTM) layers;

a plurality of conformer layers; or

a plurality of transformer layers.

9. A computer-implemented method when executed on data processing hardware causes the data processing hardware to perform operations comprising:

receiving, as input to an automatic speech recognition (ASR) model, a sequence of acoustic frames; and

performing, using the ASR model, streaming speech recognition and non-streaming speech recognition on the sequence of acoustic frames by:

generating, by a first encoder, at each of a plurality of output steps, a first higher order feature representation for a corresponding acoustic frame in the sequence of acoustic frames;

receiving, as input to a second encoder, the first higher order feature representation generated by the first encoder at each of the plurality of output steps;

generating, by the second encoder, at each of the plurality of output steps, a second higher order feature representation for a corresponding first higher order feature frame;

when the ASR model is operating in a non-streaming mode:

receiving, as input to a single decoder, the second higher order feature representation generated by the second encoder at each of the plurality of output steps; and

generating, by the single decoder, at each of the plurality of time output steps, a first probability distribution over possible speech recognition hypotheses; and

when the ASR model is operating in a streaming mode:

receiving, as input to the single decoder, the first higher order feature representation generated by the first encoder at each of the plurality of output steps; and

generating, by the single decoder, at each of the plurality of output steps, a second probability distribution over possible speech recognition hypotheses;

wherein a training process trains the model to learn how to operate in the non-streaming mode and the streaming mode by:

receiving a set of training utterances; and

jointly training, in a single stage, the ASR model using the set of training utterances by:

for each training utterance of the set of training utterances:

sampling the training utterance of the set of training utterances;

stochastically selecting one of the non-streaming mode of the ASR model and the streaming mode of the ASR model;

training the selected one of the non-streaming mode of the ASR model and the streaming mode of the ASR model; and

calculating a loss for the training utterance only once;

generating a first loss for the first encoder;

generating a second loss for the second encoder; and

minimizing a weighted sum of the first loss and the second loss.

10. The method of claim 9 , wherein the second encoder generates the second higher order feature representation without receiving any of the acoustic frames as input.

11. The method of claim 10 , wherein the operations further comprise generating partial speech recognition results based on the second probability distribution over possible speech recognition hypotheses.

12. The method of claim 10 , wherein the operations further comprise, when the ASR model is operating in the streaming mode:

receiving, as input to a prediction network, a sequence of non-blank symbols output by a final Softmax layer;

generating, by the prediction network, at each of the plurality of output steps, a dense representation;

receiving, as input to a joint network, the dense representation generated by the prediction network at each of the plurality of output steps and the first higher order feature representation generated by the first encoder at each of the plurality of output steps; and

generating, by the joint network, at each of the plurality of output steps, the second probability distribution over possible speech recognition hypotheses.

13. The method of claim 10 , wherein the operations further comprise, when the ASR model is operating in the non-streaming mode:

receiving, as input to a prediction network, a sequence of non-blank symbols output by a final Softmax layer;

generating, by the prediction network, at each of the plurality of output steps, a dense representation;

receiving, as input to a joint network, the dense representation generated by the prediction network at each of the plurality of output steps and the second higher order feature representation generated by the second encoder at each of the plurality of output steps; and

generating, by the joint network, at each of the plurality of output steps, the first probability distribution over possible speech recognition hypotheses.

14. The method of claim 9 , wherein the decoder comprises a long short-term memory (LSTM)-based prediction network.

15. The method of claim 9 , wherein the decoder comprises a V2 embedding look-up table-based prediction network.

16. The method of claim 9 , wherein the first encoder comprises a causal encoder comprising one of:

a plurality of unidirectional long short-term memory (LSTM) layers;

a plurality of conformer layers; or

a plurality of transformer layers.

17. The method of claim 9 , wherein the second encoder comprises a non-causal encoder comprising:

one or more bi-directional long short-term memory (LSTM) layers;

a plurality of conformer layers; or

a plurality of transformer layers.