A Kalman filtering noise canceler for PDC speech enhancement

A Kalman Filtering Noise Canceler for

PDC Speech EnhancementYoshinori MIKI, Hirohito SUDA and Tomoyuki OHYAR&D Department," T Mobile Communications Network Inc. I 1-2356 Take, Yokosuka-shi, Kanagawa-ken, 238-03 Japan

Abstract Speech enhancement techniques suitable for digital mobile communication systems are proposed. The proposed methods improve the speech quality and enlarge the continuous conversation time of any mobile radio system. The enhancement techniques, whose complexity is about l/lOth that of the speech codec, are easily implemented in the same chip as the speech CODEC. The core of the proposed techniques is a noise canceler based on a Kalman filter and signal-state estimation. We employ the Kalman filtering algorithm for its optimum filtering performance for time varying signals. In order to achieve effective noise cancelation with the filter, the filtering parameters are controlled according to the estimated signalstate. The speech enhancement performance of the proposed noise canceler is evaluated in terms of mean opinion score (MOS). The proposed techniques improve noisy speech quality by half a point in MOS a t the noise level of 50 dBA. Improvements in transmission time ratio by employing the proposed noise canceler in the Japanese PDC (Personal Digital Cellular) system are evaluated. The combination of voice operated transmission (VOX) in PDC and the noise canceling techniques save the battery life. The transmission period ratio is improved by 20% at the noise level of 60 dBA. We describe a speech detection scheme and processing method for reducing the speech degradation due to VOX operation.

-

I. Introduction Cellular telephone terminals are used in various environments including streets and cars. Especially in streets, significant environmental noise often surrounds the user and contaminates the speech signal and degrades the quality of the telephone conversations. Canceling the environmental noise increases the input speech SNR and improves the speech quality. Noise cancelation in a cellular telephone terminal improves not only speech quality, but also battery power consumption. In a speech transmission system employing a VOX (Voice Operated Transmission)[ 11, if it is difficult to distinguish between large environmental noise and the voice signal, the radio signal is transmitted even during non-utterance (nonspeaking) periods. Noise cancelation prevents erroneous VOX operation from draining the battery by reducing the environmental noise. In the following section, we discuss the speech transmission system of Fig. 1, which employs a noise canceler and VOX.

Mobile station

YTx

Base station

iSpeech inputII

:-"" !Control

:Speech output

Speech

CoderI

--k)

Fig.1 Speech Transmission System employing Noise Canceler and VOX0-7803-2955-4/95/$4.00

0IEEE

718

We propose a speech enhancement and VOX algorithm suitable for digital mobile communication systems. Its effects on speech quality and battery life are

evaluated as applied to the PDC (Personal Digital Cellular) system[2][3].

Table 1 Description of the Signal-StateSignalstate Noise canceling Description of the signal-stae Silent noise, Stable utterance Silent noise, Start or end of utterance Silent noise, Non-utterance Environmental noise, Non-utterance Environmental noise, Start or end of utterance Environmental noise, Stable utterance

11 10 0 20 2122

Off

offOff On (strong)

11. Noise Canceler A . Noise cancelation algorithm We propose a noise canceling scheme based on Kalman filtering[4] because of its optimum performance with time varying signals. Fig. 2 shows the proposed block diagram of the noise canceler. The input signal-state, which is explained in Table 1, is estimated in order to control the parameters of the Kalman filter. In order to achieve effective noise cancelation with Kalman filtering, the filtering parameters, which are statistical values of the speech signal and environmental noise, are controlled according to the estimated input signal-state. The signal-state is determined every 40ms (one frame) and the signal-state transitions are defined as shown in Fig. 3. The average signal power in each frame, the signal power difference between current and previous frames, the signal spectrum envelope, and its difference are used to determine the state transitions, and also, speech detections. From state 20 to 22, the Kalman filter is applied to the input signal. For the other states, however, the output signal is the same as the input. At state 20, which represents a nonutterance frame in a noisy environment, the parameter for the Kalman filter is controlled to strongly reduce the input noise. At state 21, which represents the start or the end of an utterance frame in a noisy environment, the parameter for the Kalman filter is controlled to weakly reduce the input noise.

On(weak) On

Fig.3 State-transition Definition This is because avoiding speech signal suppression at the start or the end of an utterance frame is particularly important for assuring the intelligibility of conversations. At state 22, which represents an utterance frame in a noisy environment, the parameter for the Kalman filter is controlled to moderately reduce the input noise. In the Kalman filter, we employ a second order autoregressive (AR) model for the speech (source) signal and the environmental (additional) noise. The AR model parameters consist of two linear predictive coefficients (LPC) and the variance of the innovation process. For strongly reducing the input noise, we enlarge the variance of the innovation process of the noise more than is observed. For weakly reducing the input noise, we enlarge the variance of the signal more than is observed. In order to reduce the computational complexity, filtering coefficient calculations are performed only three times (sam …… 此处隐藏：12043字，全部文档内容请下载后查看。喜欢就下载吧 ……