Space-time voice activity detection for mobile environments

다국어 초록 (Multilingual Abstract)

Recently, mobile devices such as cellular phones, personal digital assistants are frequently used in real life for their convenience. However, the interface between the mobile devices and a human being is widely used in button based methods or a touch sensor technology. While speech-based interfaces are known for their naturalness and convenience, the technical difficulties, such as speech recognition in the noisy environment, impede the development of the speech interface. The performance degradation of speech recognizers in noisy environments can be largely compensated by the effective voice activity detection (VAD) algorithms.
VAD is the problem of distinguishing the human speech from the other sounds. It can play an important role in mobile conditions which have a low power capacity and always run for the unexpected input. By using a well-operated VAD module in the mobile environments, system power can be reduced by ignoring the noise signals and the system properly executes a desired action even though voice signal is received on a sudden. When the speech-based interfaces are used in the mobile environment, noise signals other than the user's voice must be rejected by the systems as unexpected signals. Because the conventional voice activity detection only focuses on discriminating the voice and noise, it is necessary to take different approach.
To reject the unexpected signals originated from noise sources; this thesis proposes a new algorithm that can detect user's voice in space-time domain using a microphone array. The proposed method uses the sound source localization to detect target signals in 3-dimensional space, and improves the performance of the end-point voice activity detection in time domain by using the reduced background noise. The main issues in mobile environment voice activity detection are signal-to-noise ratio (SNR), reverberation time (RT), the angles of the incident noise, and the noise types. Corresponding experiments were conducted for these factors. The experimental results show that the proposed algorithm reduces the voice activity detection error rate by 32.3% relative to the conventional method in the PDA environment and by 27.3% relatively in robot environment for both stationary noises and non-stationary noises.

목차 (Table of Contents)

• Chapter 1 Introduction = 1
• Chapter 2 Related Works = 4
• 2-1. Speech and Speaker Recognition = 4
• 2-2. Voice Activity Detection = 6
• Chapter 3 Space-Time Voice Activity Detection = 8
• 3-1. Energy based Detection = 10
• 3-2. Sound Source Localization = 11
• 3-2-1. Steered Response Power-Phase Transform = 12
• 3-2-2. Virtual Relocation of Microphone = 13
• 3-2-3. Decision Strategy for Noisy Signals = 16
• 3-3. Speech Enhancement = 17
• 3-4. Time Domain End-Point Detection = 19
• 3-5. Summary = 21
• Chapter 4 Experiments = 22
• 4-1. PDA Environment Experiments = 22
• 4-1-1. Simulation Data Evaluation = 22
• 4-1-2. Real Data Evaluation = 30
• 4-2. Robot Environment Experiments = 37
• Chapter 5 Conclusion = 44
• References = 46