Utterance Confidence Measure for End-to-End Speech Recognition with Applications to Distributed Speech Recognition Scenarios

Neural Utterance Confidence Measure for RNN-Transducers and Two Pass Models

“…[8] shows that acoustic embedding plays an important role for measuring confidence score. Although joint network uses these features for predictions but [14] show that explicitly passing these features to NCM Model gives significant boost to the model performance.…”

“…Several works train neural network for this binary classification task [10,13]. Recently, [14] model [15] to train a neural classifier. We extend this work to RNN-T and 2-pass frameworks.…”

“…We extend this work to RNN-T and 2-pass frameworks. The later framework uses a streaming model such as RNN-T and an offline decoder, both of which behave differently from monotonic attention-based ASR model used in [14] 3. END-TO-END SPEECH RECOGNITION…”

“…We extend the techniques discussed in [14] for calculating neural confidence measure (NCM) for RNN-T and 2-pass models. In [14] authors compute the confidence score for the prediction of Monotonic Chunkwise Attention (MoChA) [1,15] decoder which is attention-based decoder. Models using RNN-Transducer framework are trained with a different objective function, which results in decoder behaviour as well as features different form those of attention-based models.…”

“…2-pass models provide a richer set of features which can be helpful for predictor feature-based confidence score models as shown in [14]. We use different combination of features from transcription network, prediction network and joint network for confidence score on the prediction of RNN-T models.…”