Abstract: Using an encoder neural network model, an encoder vector is computed, the encoder vector comprising a vector representation of a current portion of input data in an input sequence. Using a prediction neural network model, a prediction vector is predicted, the prediction performed using a previous prediction vector and a previous output symbol corresponding to a previous portion of input data in the input sequence. Using a joint neural network model, a joint vector corresponding to the encoder vector and the prediction vector is computed, the joint vector multiplicatively combining each element of the encoder vector with a corresponding element of the prediction vector. Using a softmax function, the joint vector is converted to a probability distribution comprising a probability that a current output symbol corresponds to the current portion of input data in the input sequence.

Abstract: An embodiment for speech-to-text auto-scaling of computational resources is provided. The embodiment may include computing a delta for each word in a transcript between a wall clock time and a time when the word is delivered to a client. The embodiment may also include submitting the deltas to a group of metrics servers. The embodiment may further include requesting from the group of metrics servers current values of the deltas. The embodiment may also include determining whether the current values of the deltas exceed a pre-defined max-latency threshold. The embodiment may further include adjusting the allocated computational resources based on a frequency of the current values of the deltas that exceed the pre-defined max-latency threshold. The embodiment may also include creating a histogram from the current values of the deltas and scaling-up the allocated computational resources based on a percentage of data points that fall above the pre-defined max-latency threshold.

Abstract: One or more audio data is received. An expected bitrate of the one or more audio data is determined. An input bitrate of the one or more audio data is determined. An R value using the expected bitrate and the input bitrate is determined. The R value is compared to an R threshold.

Abstract: One or more audio data is received. An expected bitrate of the one or more audio data is determined. An input bitrate of the one or more audio data is determined. An R value using the expected bitrate and the input bitrate is determined. The R value is compared to an R threshold.

Abstract: An embodiment for speech-to-text auto-scaling of computational resources is provided. The embodiment may include computing a delta for each word in a transcript between a wall clock time and a time when the word is delivered to a client. The embodiment may also include submitting the deltas to a group of metrics servers. The embodiment may further include requesting from the group of metrics servers current values of the deltas. The embodiment may also include determining whether the current values of the deltas exceed a pre-defined max-latency threshold. The embodiment may further include adjusting the allocated computational resources based on a frequency of the current values of the deltas that exceed the pre-defined max-latency threshold. The embodiment may also include creating a histogram from the current values of the deltas and scaling-up the allocated computational resources based on a percentage of data points that fall above the pre-defined max-latency threshold.

Abstract: Using an encoder neural network model, an encoder vector is computed, the encoder vector comprising a vector representation of a current portion of input data in an input sequence. Using a prediction neural network model, a prediction vector is predicted, the prediction performed using a previous prediction vector and a previous output symbol corresponding to a previous portion of input data in the input sequence. Using a joint neural network model, a joint vector corresponding to the encoder vector and the prediction vector is computed, the joint vector multiplicatively combining each element of the encoder vector with a corresponding element of the prediction vector. Using a softmax function, the joint vector is converted to a probability distribution comprising a probability that a current output symbol corresponds to the current portion of input data in the input sequence.