Quality estimation for automatic speech recognition
Methods and systems are provided for implementing quality estimation for automatic speech recognition, and more specifically training an ASR model, and training a QE model to perform word error rate prediction upon the trained ASR model. The ASR model may be a transformer learning model having an architecture including an encoder including multi-head attention layers, and a memory encoder including a masking multi-head attention layer. The QE model may include a binary classification model and a regression model, where the binary classification model is based on a discrete statistical distribution, and the regression model is based on a continuous statistical distribution. Training the ASR model may produce output having variable word error rates, and the QE model may be trained based on empirical word error rates of the ASR model. The QE model may predict performance of the ASR model without labor-intensive labeling to generate ground truth.
1 . A method comprising:
training a transformer learning model on inputs comprising a sequence of audio tokens, the trained transformer learning model being executable by one or more processors of a computing system to output a sequence of feature representations; and
training a quality estimation learning model on inputs comprising the sequence of feature representations, the trained quality estimation learning model being executable by the one or more processors to output a probability of a word error rate value.
2 . The method of claim 1 , wherein the transformer learning model comprises:
an encoder trained to be executable by the one or more processors to output a plurality of speech features; and
a memory encoder trained on a feature matrix comprising an embedding of the plurality of speech features.
3 . The method of claim 2 , wherein the memory encoder comprises a masking multi-head attention layer; and
further comprising:
masking, at the masking multi-head attention layer, a plurality of features of the feature matrix; and
computing attention among non-masked features of the feature matrix.
4 . The method of claim 1 , wherein the quality estimation learning model comprises a binary classification model and a regression model.
5 . The method of claim 4 , wherein the binary classification model is trained to be executable by the one or more processors to output one of:
a perfect classification comprising a probability of a word error rate value of zero, and
a non-perfect classification.
6 . The method of claim 4 , wherein the binary classification model is trained to be executable by the one or more processors to perform binary classification based on a Bernoulli statistical distribution.
7 . The method of claim 4 , wherein the regression model is trained to be executable by the one or more processors to output a probability of a word error rate value of non-zero, based on a beta statistical distribution.
8 . A system comprising:
one or more processors; and
memory communicatively coupled to the one or more processors, the memory storing computer-executable modules executable by the one or more processors that, when executed by the one or more processors, perform associated operations, the computer-executable modules comprising:
a transformer learning model module trained to be executable by the one or more processors to receive inputs comprising a sequence of audio tokens, and output a sequence of feature representations; and
a quality estimation learning model trained to be executable by the one or more processors to receive inputs comprising the sequence of feature representations, and output a probability of a word error rate value.
9 . The system of claim 8 , wherein the transformer learning model module comprises:
an encoder module trained to be executable by the one or more processors to output a plurality of speech features; and
a memory encoder module trained on a feature matrix comprising an embedding of the plurality of speech features.
10 . The system of claim 9 , wherein the memory encoder module comprises a masking multi-head attention layer module;
wherein the masking multi-head attention layer module is executable by the one or more processors to mask a plurality of features of the feature matrix; and
wherein the masking multi-head attention layer module is further executable by the one or more processors to compute attention among non-masked features of the feature matrix.
11 . The system of claim 9 , wherein the quality estimation learning model module comprises a binary classification model module and a regression model module.
12 . The system of claim 11 , wherein the binary classification model module is trained to be executable by the one or more processors to output one of:
a perfect classification comprising a probability of a word error rate value of zero, and
a non-perfect classification.
13 . The system of claim 11 , wherein the binary classification model module is trained to be executable by the one or more processors to perform binary classification based on a Bernoulli statistical distribution.
14 . The system of claim 11 , wherein the regression model module is trained to be executable by the one or more processors to output a probability of a word error rate value of non-zero, based on a beta statistical distribution.
15 . A computer-readable storage medium storing computer-readable instructions executable by one or more processors, that when executed by the one or more processors, cause the one or more processors to perform operations comprising:
training a transformer learning model on inputs comprising a sequence of audio tokens, the trained transformer learning model being executable by the one or more processors to output a sequence of feature representations; and
training a quality estimation learning model on inputs comprising the sequence of feature representations, the trained quality estimation learning model being executable by the one or more processors to output a probability of a word error rate value.
16 . The computer-readable storage medium of claim 15 , wherein the transformer learning model comprises:
an encoder trained to be executable by the one or more processors to output a plurality of speech features; and
a memory encoder trained on a feature matrix comprising an embedding of the plurality of speech features.
17 . The computer-readable storage medium of claim 16 , wherein the memory encoder comprises a masking multi-head attention layer; and
wherein the operations further comprise:
masking, at the masking multi-head attention layer, a plurality of features of the feature matrix; and
computing attention among non-masked features of the feature matrix.
18 . The computer-readable storage medium of claim 15 , wherein the quality estimation learning model comprises a binary classification model and a regression model.
19 . The computer-readable storage medium of claim 18 , wherein the binary classification model is trained to be executable by the one or more processors to output one of:
a perfect classification comprising a probability of a word error rate value of zero, and
a non-perfect classification.
20 . The computer-readable storage medium of claim 18 , wherein the regression model is trained to be executable by the one or more processors to output a probability of a word error rate value of non-zero, based on a beta statistical distribution.