How Voice Assistants Distinguish Speakers After a Wakeword

This patent describes a method for voice assistants to better understand who is speaking after a 'wakeword' by comparing voice characteristics to improve speech recognition accuracy, especially in multi-speaker environments.

What it covers

This patent outlines a computer-implemented method for automatic speech recognition that focuses on distinguishing between speakers after a 'wakeword' is detected. First, the system receives audio data and processes it to identify a 'wakeword' spoken by a 'first speaker' (Claim 1). It then takes subsequent portions of audio and compares their characteristics to those of the original wakeword. For example, if a second person speaks, the system determines a 'first similarity data' between the wakeword and that second person's speech (Claim 1). It then calculates a 'first weight' based on this similarity. If the original speaker continues, the system determines 'second similarity data' between the wakeword and the original speaker's continued speech, calculating a 'second weight' that is greater than the first (Claim 1). By comparing these weighted similarity measures, the system determines that the speech from the original speaker is 'more likely' to have come from the person who spoke the wakeword. Finally, it processes the speech from the original speaker to convert it into text and execute a command (Claim 1).

What it doesn't cover

• —Does not cover speech recognition systems that do not rely on a 'wakeword' to initiate the process.
• —Does not cover general speaker identification or diarization that is not specifically tied to processing subsequent speech after a wakeword.
• —Does not cover methods that do not use similarity data and calculated weights to distinguish between speakers after the wakeword.
• —Does not cover systems where the 'second weight' is not determined to be greater than the 'first weight' when comparing the original speaker's subsequent speech to another speaker's speech.

The clever bit

The novelty lies in using the initial 'wakeword' spoken by a specific person as a reference point. The system then generates 'similarity data' and 'weights' by comparing subsequent speech segments to the wakeword's unique audio characteristics. These weights help the system prioritize and accurately process commands from the intended speaker, even when other voices are present.

Why it matters

This technology is crucial for improving the accuracy and user experience of voice-controlled devices, especially in homes with multiple people. By precisely distinguishing who is speaking after a command, it helps voice assistants like Amazon Alexa or Google Assistant respond correctly to the intended user. This reduces errors and makes smart speakers more reliable and useful in everyday life.

Real-world examples

• 1.Amazon Echo devices
• 2.Google Nest Hub
• 3.Smart speakers with voice assistants
• 4.Voice-controlled smart home devices