Efficient memory transformer based acoustic model for low latency streaming speech recognition
In one embodiment, a method includes accessing a machine-learning model configured to generate an encoding for an utterance by using a module to process data associated with each segment of the utterance in a series of iterations, performing operations associated with an i-th segment during an n-th iteration by the module, which include receiving an input comprising input contextual embeddings generated for the i-th segment in a preceding iteration and a memory bank storing memory vectors generated in the preceding iteration for segments preceding the i-th segment, generating attention outputs and a memory vector based on keys, values, and queries generated using the input, and generating output contextual embeddings for the i-th segment based on the attention outputs, providing the memory vector to the module for performing operations associated with the i-th segment in a next iteration, and performing speech recognition by decoding the encoding of the utterance.
1. A method comprising, by one or more computing systems:
accessing a machine-learning model configured to generate an encoding for an utterance by using a module to process, in a series of hierarchical layers associated with the module, data associated with each of a sequence of segments of the utterance;
performing, by the module, operations associated with an i-th segment during an n-th iteration layer of the series of hierarchical layers, comprising:
receiving an input comprising (1) input contextual embeddings generated for the i-th segment in a preceding lower layer of the series of hierarchical layers and (2) a memory bank storing memory vectors generated in the preceding lower layer for one or more of the segments preceding the i-th segment, wherein the one or more of the segments exclude the i-th segment, and wherein each memory vector is determined based on an attention operation based on each of the one or more of the segments preceding the i-th segment;
generating attention outputs and a memory vector based on keys, values, and queries generated using the input; and
generating, based on the attention outputs, output contextual embeddings for the i-th segment;
providing the memory vector to the module for performing operations associated with the i-th segment in a next higher layer of the series of hierarchical layers; and
performing speech recognition by decoding the encoding of the utterance.
2. The method of claim 1 , wherein the input contextual embeddings for the i-th segment during the n-th layer comprises an embedding representing the i-th segment generated in the preceding lower layer and an embedding representing a right context block associated with the i-th segment generated in the preceding lower layer.
3. The method of claim 2 , further comprising:
making a copy of the embedding representing the right context block; and
putting the copy at a beginning of the received input.
4. The method of claim 1 , wherein generating the output contextual embeddings is based on a point-wise feed-forward network.
5. The method of claim 1 , wherein the output contextual embeddings for the i-th segment during the n-th layer comprises an embedding representing the i-th segment generated in the n-th layer and an embedding representing a right context block associated with the i-th segment generated in the n-th layer.
6. The method of claim 1 , wherein the output contextual embeddings for the i-th segment during the n-th layer are input contextual embeddings for the i-th segment during the next higher layer.
7. The method of claim 1 , generating the encoding for the utterance comprising:
after the series of hierarchical layers, identifying, from output contextual embeddings for each segment of the utterance, an embedding representing the segment; and
combining the identified embeddings of the segments of the utterance.
8. The method of claim 1 , further comprising:
after the n-th layer, updating the memory bank by adding the memory vector generated during the n-th layer.
9. The method of claim 1 , further comprising:
generating a summary vector for the i-th segment during the n-th layer, wherein the summary vector is based on a mean of the i-th segment; and
assigning a zero attention-weight between the summary vector and the memory bank.
10. The method of claim 1 , wherein the keys comprise at least a key carried over from one or more segments preceding the i-th segment.
11. The method of claim 1 , wherein the values comprise at least a value carried over from one or more segments preceding the i-th segment.
12. One or more computer-readable non-transitory storage media embodying software that is operable when executed to:
access a machine-learning model configured to generate an encoding for an utterance by using a module to process, in a series of hierarchical layers associated with the module, data associated with each of a sequence of segments of the utterance;
perform, by the module, operations associated with an i-th segment during an n-th layer of the series of hierarchical layers, comprising:
receiving an input comprising (1) input contextual embeddings generated for the i-th segment in a preceding lower layer of the series of hierarchical layers and (2) a memory bank storing memory vectors generated in the preceding lower layer for one or more of the segments preceding the i-th segment, wherein the one or more of the segments exclude the i-th segment, and wherein each memory vector is determined based on an attention operation based on each of the one or more of the segments preceding the i-th segment;
generating attention outputs and a memory vector based on keys, values, and queries generated using the input; and
generating, based on the attention outputs, output contextual embeddings for the i-th segment;
provide the memory vector to the module for performing operations associated with the i-th segment in a next higher layer of the series of hierarchical layers; and
perform speech recognition by decoding the encoding of the utterance.
13. The media of claim 12 , wherein the input contextual embeddings for the i-th segment during the n-th layer comprises an embedding representing the i-th segment generated in the preceding lower layer and an embedding representing a right context block associated with the i-th segment generated in the preceding lower layer.
14. The media of claim 13 , wherein the software is further operable when executed to:
make a copy of the embedding representing the right context block; and
put the copy at a beginning of the received input.
15. The media of claim 12 , wherein generating the output contextual embeddings is based on a point-wise feed-forward network.
16. The media of claim 12 , wherein the output contextual embeddings for the i-th segment during the n-th layer comprises an embedding representing the i-th segment generated in the n-th layer and an embedding representing a right context block associated with the i-th segment generated in the n-th layer.
17. A system comprising: one or more processors; and a non-transitory memory coupled to the processors comprising instructions executable by the processors, the processors operable when executing the instructions to:
access a machine-learning model configured to generate an encoding for an utterance by using a module to process, in a series of hierarchical layers associated with the module, data associated with each of a sequence of segments of the utterance;
perform, by the module, operations associated with an i-th segment during an n-th layer of the series of hierarchical layers, comprising:
receiving an input comprising (1) input contextual embeddings generated for the i-th segment in a preceding lower layer of the series of hierarchical layers and (2) a memory bank storing memory vectors generated in the preceding lower layer for one or more of the segments preceding the i-th segment, wherein the one or more of the segments exclude the i-th segment, and wherein each memory vector is determined based on an attention operation based on each of the one or more of the segments preceding the i-th segment;
generating attention outputs and a memory vector based on keys, values, and queries generated using the input; and
generating, based on the attention outputs, output contextual embeddings for the i-th segment;
provide the memory vector to the module for performing operations associated with the i-th segment in a next higher layer of the series of hierarchical layers; and
perform speech recognition by decoding the encoding of the utterance.
18. The system of claim 17 , wherein the input contextual embeddings for the i-th segment during the n-th layer comprises an embedding representing the i-th segment generated in the preceding lower layer and an embedding representing a right context block associated with the i-th segment generated in the preceding lower layer.
19. The system of claim 18 , wherein the processors are further operable when executing the instructions to:
make a copy of the embedding representing the right context block; and
put the copy at a beginning of the received input.
20. The system of claim 17 , wherein generating the output contextual embeddings is based on a point-wise feed-forward network.