Confidence checking for speech processing and query answering

This invention describes a way for a speech recognition system to check its own work. It uses a special type of artificial intelligence called a recurrent neural network to analyze different parts of the speech signal, like individual sound frames and phone sounds, to create a unique digital fingerprint for the recognized words. This fingerprint is then used to confirm if the speech recognition result was correct, or to give it a confidence score, especially for the highest-scoring word sequence and when a 'wakeword' is detected.