Abstract: In a computer-implemented voice recognition method, an audio signal is provided to an acoustic model, and processed to determine phonemes from the audio signal The phonemes are provided to a decoding graph having a plurality of search paths corresponding to sequences of phonemes. An output from the decoding graph is determined according to one of the plurality of search paths, in which the plurality of search paths in the decoding graph includes a plurality of command search paths defining a plurality of sequences of phonemes for a respective plurality of voice commands and a watermark search path defining a sequence of phonemes for a watermark of the decoding graph.

Abstract: A method for processing and responding to audio input signals includes receiving, at a first device, an audio input signal having a trigger prompt. At the first device, the audio input signal is processed to determine a first quality metric of the audio input signal. The first device transmits the first quality metric over a local network and monitors whether a higher second quality metric of the audio input signal is transmitted by a second device over the local network. If a higher second quality metric is not received by the first device, by the first device responds to the trigger prompt.