There is a method and system for translating speech to text, the speech having been received by a client device. A user utterance corresponding to the speech is received. A first predicted text corresponding to the user utterance and a first confidence score corresponding to the first predicted text are determined using a local graph. The user utterance is transmitted to a server. A second predicted text corresponding to the user utterance and a second confidence score corresponding to the second predicted text is received from the server. If the first confidence score is higher than the second confidence score, the first predicted text is output.