Endpoint Detection（ 端点侦测）：端点检测（端点侦测）.ppt

-*- Endpoint Detection( 端點偵測) Jyh-Shing Roger Jang (張智星) /jang MIR Lab, CSIE Dept National Taiwan Univ., Taiwan Intro. To Endpoint Detection Endpoint detection (EPD,端點偵測) Goal: Determine the start and end of voice activity Also known as voice activity detection (VAD) Importance Acts as preprocessing for many recognition tasks Requires as small computing power as possible Operation scenarios for speech recognition Off-line for “pushing to talk” On-line for “continuously listening” Two Types of Approaches to EPD Time-domain methods Volume ZCR (zero crossing rate) HOD (high-order difference) Frequency-domain methods Variance of spectrum Entropy of spectrum Typical Frameworks to EPD Thresholding Simple thresholding Compute a features (e.g., volume) from each frame Select a threshold vth Any frame with a volume higher than vth is considered positive Combined thresholding Use two features (e.g., volume and ZCR) to have more complicated decision making Classification Take more than one features Perform binary classification Negative?sil or noise Positive?sound activity Sequence alignment Use hidden Markov models (HMM) for sequence alignment Performance Evaluation for EPD Types of errors False rejection positive ? negative False acceptance negative ? positive Performance evaluation Start end position accuracy Frame-based accuracy EPD by Volume Only The simplest method for EPD Four intuitive way to select vth? vth = vmax*a vth = vmedian*b vth = vmin*g vth = v1*d EPD by Volume Only (II) Unfortunately… Most of the thresholds fail one way or another. Dataset-based fine-tuning of a, b, g, d is always advisable. Under what situations do they fail? Plosive sounds Silence too long Total-zero frame Unstable frame 1 EPD by Volume Only (III) A presumably better way to select vth vlower = 3rd percentile of all ascending volume vupper = 97th percentile of all ascending volume vth = (vupper-vlower)*k+vlower Why do we need to use percentile? To deal with plosive sounds To deal