The research prertnted here investigates the applicatiun qf the Narruivbund Adaptive. Multi-Rate (NB-AMR) speech codec and the Wideband. AMR (NIB-AMR) ...
Range Extension and Short Range Performance Enhancement in TDMA Digital Cellular" Bo Wei
Jerry D. Gibson
Department of Electrical Engineering Southem Methodist University Dallas, TX 75275
Media Arts Technology Program and Electrical & Computer Engineering University of California, Santa Barbara Santa Barbara, CA 93 106-6065 speech codec that has several speech coder rates and that includes both the IS-641 and the US1 speech codecs as submodes 121. More recent work by the 3GPP (3rd Generation Partnership Project) has led to the standardization of a wideband AMR (WB-AMA) codec for encoding wideband speech (SO HZ to 7 kHz) at rates from 6.6 up to 23.85 kbps. We propose an enhancement to T M I A 136 that provides improvements both in clear channel speech quality and in range extension for fading channels by incorporating the NB and WB-AMR codecs in the d 4 DQPSK modulation format with adaptive channel coding rates. When the channel degrades, lower bit rate speech codecs are used to release additional bits for error protection and hence maintain robustness, but when the channel improves, higher bit rate speech codecs can be selected to provide better speech quality, since fewer bits are needed for error ConUoI coding. We are able to provide range extension below a CA of IS dB while also improving perfomlance in the existing operational range above 1.5 dB. We keep the time slot formats unchanged so that our method is completely compatible with existing 136 systems. We briefly introduce the original TDMA system, emphasizing its speech and channel coding methods in Section 2. The introduction of NB-AMR and WB-AMR is given in Section 3. In Section 4,a detailed description of our new TDMNAMR approach is provided. Performance evaluation and system realization issues are given in Section 5 . Conclusions are in Section 6.
Abstract A new pstem enhancement method is proposed for the
EIAflIA-I36 system ofering both channel operational range extension and iniproved perfornrance within the current operational range. The research prertnted here investigates the applicatiun qf the Narruivbund Adaptive Multi-Rate (NB-AMR) speech codec and the Wideband AMR (NIB-AMR) codec, both originally designed for the 200kHz GSM clionnel, in the TD.M (TL4LEIA-1345) 30 kUz sy.ytem. In particular, we investigate adaptiwly allocating bits hefiveen NBM'B speech coding and error control coding within the liniited choiinel banuividth. Four modes out of seventeen have been carefirlly chosen for the new TD.WAMR system. Switching befiveen codec rates as channel conditions change produces runge extension below a C/f of 15 dB while rrlso improving perfonnonce in the existing operational rwige above I5 dB. Tinre slot farm at.^ are unchanged so that our method is conqdetely compatible with existing I36 s.v.stems.
1 Introduction The existing TDMA (136) speech codec, the IS-641 Enhanced Full Rate (EFR) vocoder, operates at a fixed bit rate which is 7.4 kbps and does not allow the reallocation of bits to channel error protection as channel conditions degrade. There has been recent interest in improving the voice quality under both clear channel and fading channel conditions [I, 21. Some improvements for the TIAEIA 136 system, designated l56+ considered using 8-PSK modulation and the US1 speech coder that operates at 12.2 kbps 121. Other efforts have involved the European Telecommunications Standard Institute (ETSI) program to develop a narrowband adaptive muftirate (NB-AMR)
2 Speech and Channel Coding in TDMA-136 System IS-641, the EFR voice codec standard of 136, uses algebraic CELP (ACELP) to reduce the effects of a large codebook on storage and search complexity and improves the quality of the reconstructed speech over the original 7.9.5 kbps VSELP codec. The encoder produces 148 bits
'lhis research usassupponed by Nokia Mobile Phones. [rving, Texas. The authors &"fully acknowledge the guidance of Paul Meche and Steve WLllhoRduring this w o k
0-7803-7576-9!02$17.00 Q 2002 IEEE
553
Project) in 2001. The WB-AMR codec is intended not only for the existing GSM network, but also for future mobile radio systems of the Third Generation. The WBAMR codec has 9 modes (specifically, 6.60, 8.85, 12.65, 14.25, 15.85, 18.25, 19.85, 23.05 and 23.85 kbps) 161 where the sampling rate is 16000 Hz and processing is performed on 20 ms (320 sample) frames. Wideband speech coding improves the speech quality by not only including high-frequencies but also by low-frequency extension.
every 20 ms, resulting in a bit rate of 7.4 kbps. As the usable data rate on a n/4 DQPSK Digital Trafiic Channel (DTC) is 13 kbps, this leaves 5.6 kbps for channel coding. The 148 bits per 20 ms from the speech encoder are subjectively ordered and divided into three classes. There are 48 class I A bits, 48 class 1B bits and 52 class 2 bits. A 7-bit CRC code is used with the class 1A bits lo ensure these most perceptually significant hits are correctly received at the channel decoder. If there is an error, a Bad Frame Indicator (BFI) is set and an error concealment procedure is applied to this frame. The 96 class 1 bits and the 7 CRC bits are protected by a rate % convolutional code with memory order 5. Now the total encoded bits are 216 including the 5 tail bits. The 216-bit code is punctured by removing 8 bits. The remaining 52 class 2 bits that are not protected are combined together uith the 208 protected bits. A total of 260 bits are interleaved across two time slots and inserted into the DATA field of the DTC, where an example of one TDMA frame structure in a 3 0 - H RF ~ channel is shown in Figure 1.
-
4 Application of N B N B Hybrid A M R Speech Codec in 30 kHz Channel In this work, we investigate the potential application of switchable NE-AMR and WB-AMR to the 30 kHz n/4 DQPSK 136 carriers. Although NB-AMR has been adopted by GSM and leads to very attractive performance enhancements, it has not been applied in North American TDMA systems with 30 kI4z channels., The difficulty is that the TIAEIA-136 30 kHz carrier provides much fewer bits (244 bits) for the 20 ms coded speech frame and for channel coding than the GSM 200 kI3z carrier does (448 bits). Therefore, with the same AMR speech codec, bits available to be allocated for channel coding is highly limited. Alternatively, the new standardized WB-AMR also provides several low-bit-rate wideband speech codec candidates that have never been applied to any existing wireless communication system. It is also desired to take advantage of wideband speech modes to improve naturalness, presence, and intelligibility, especially for female speakers. With these considerations in mind, we investigate an adaptive multi-rate 30 kHz camer TDMA system that utilizes a switchable N B N B hybrid system. Because there are only 2 in-band bits available in each kame for mode adaptation, we need to select a maximum of 4 N B N B modes from 17 candidates (8 NB-AMR modes + 9 WB-AtVR modes) for our new system, which we call the TDMAIAMR system. A codec rate selection procedure is conducted based upon objective and subjective criteria
Thm full-rateusenorsir halfialcuwrs
14
6
65
6
85
5
1
\ / 130 symbols = 260 blls Figure 1. Example EFR frame and time slot structure
3 Adaptive Multi-Rate Speech Codec The basic idea of AMR is to provide several choices in coded speech bit rate while maintaining high quality. The narrowband AMR speech coder was standardized first by ETSI (European TelecommunicationsStandards Institute) for GSM in 1998 151. The NB-AMR speech codec operates at bit rates between 4.75 and 12.2 kbps (specifically, 12.2, 10.2, 7.95, 7.40, 6.70, 5.90, 5.15 and 4.75 kbps) and uses the principle of ACELP for all specified bit rates. The sampling rate is 8000 liz and processing is performed on 20 ms (160 sample) frames. Some of the NB-AMR speech coding rates match speech codecs specified for other standards, e.g. NB-AMR at 6.7 kbps is the ACELP codec specified in PDC-EFR, the 7.4 kbps mode is equivalent to IS-641, and 12.2 kbps mode is same as the GSM EFR codec. Wideband AMR is another AMR speech coder standard introduced by 3GPP (3rd Generation Partnership
4.1
Bit Allocation Budget and Initial Mode Selection
From Section I, we know that the TDMA 30 kHz canier provides 130 symbols for a 20 ms speech frame, which is equivalent to 260 bits uith DQPSK modulation. Since our new AMR system is required to be totally compatibleuith existing 136 systems, we must keep the time slot formats unchanged. For AMR link adaptation, the Base Station
554
(9s)and the Mobile Station (MS) need to exchange inband mode information that is a 2-bit sequence indicating 4 different modes. Applying block coding to these 2 bits results in a total of 16 in-band bits. The total number of bits available for speech coding and channel coding is now only 260-16=244 bits. By considering the channel limitation, any codec mode above 10.0 kbps is not usable because there will not be enough bit rate for error protection. The remaining modes for NB-AMR are 4.75 kbps, 5.15 kbps, 5.9 kbps, 6.7 kbps, 7.4 kbps (EFR) and 7.95 kbps, while only 6.6 kbps and 8.85 kbps are left for WB-AMR. Subjective listening tests indicate that the 4.75 kbps mode has considerable artifacts and the 6.7 kbps has performance similar to the 5.9 khps mode. Similar comments apply to the 7.4 khps and 7.95 kbps modes. With these considerations in mind, we reduce our candidate rates to basically five modes: NB5.15 kbps, NB5.9 kbps, NB7.4 khps, WB6.6 khps and WB8.85 kbps. A more sophisticated rate-cutting procedure that depends on the channel error robustness is described in the next several subsections. 4.2
CLASSZ: Fixed codebook index for 1st to 4th subfiames.
The proposed channel encoding procedure is shown in Figure 2. We applied sol? decision Viterbi decoding instead of hard decisions in this work. The receiver estimates the transmitted symbol and has basically hvo information parameters a\,ailable to it, i. e. the value of the symbol (0 or I ) and the reliability of the decision. From our experiments, NB5.15 outperforms the other four niodes in error correction because it has much more room availahle to fit in a rate 113 convolutional code, while others are using a rate 112 convolutional code.
6-- 4 r ,
4 F
Unequal Error Protection (UEP) with SoftDecision Decoding and RSCP Codes
-a 2m
124 .*md.
ou
The varying bit error sensitivity of the speech encoder output bits leads to a difference in their importance for speech reconstruction at the receiver side: Therefore, unequal error protection is highly desired in our AMR system. The target rates after channel encoding can he obtained by puncturing. In this study, w'e used recursive systematic convolutional (RSC) codes with constraint length 5 or 6 (8.85 khps only) for convolutional channel coding. Source encoder output bits are subjectively ordered first according to the 3GPP document for NBAMR 141 and W B - M R 171. Ordered hits then pass through a classification procedure where these bits are separated as Class la, Class Ib and Class 2 groups. We directly borrow the bit classification results of NB-AMR for GSM [4]. In the WE-AMR fiame structure [8], the speech encoder output hits are divided into h+'o classes, Class A and Class B. This class division method is insufficient for our application here because there are still some important bits in Class B that need to be protected. We reclassify the hits according to some hasic rules similar to the IS-641 and the NB-AMR standard:
Itlbn
Figure 2. Channel encoding block diagram for TDMAIAMR system
4.3 Error concealment We perform the error concealment procedure after channel decoding according to the NB-AMR [9] and WBAMR [IO] standards. The CRC checksum is used hy the receiver to determine if a decoded frame is "good" or not. If Class l a bits do not satisfy the 7-bit (NB-AMR)or 8-bit CRC (WB-AMR), the frame is treated as erroneousilost speech frames (BFI=l) and error concealment is applied. A State Counter monitors the number of consecutive bad frames, so the whole muting and concealment procedure is controlled by the BFI and State. Current frame coefficients are modified according to the state value or replaced from the previous good frame. To achieve better overall system enhancement, joint source-channel coding [3] can also be applied here as an error concealment procedure hy exploiting source apriori information and channel bit reliability information.
CLASS /a: Index of 1st and 2nd LSF subvectors; Most bits of adaptive codebook index for 1st and 3rd subframes; most bits of adaptive codebook gain for 1st to 4th subframes. CLASS Ib: All the rest of the hits of adaptive codebook index and adaptive codebook gain for 1st to 4th subframes; fixed codebook sign for 1st to 4th suhframes.
5 Performance Evaluation and System Realization Issues 5.1 Simulation Testbed
555
Figure 4. During uplink, BS keeps detecting the channel condition and decides which mode should be applied for the next speech frame and passes this information as Codec Mode Command (CMC) to MS in the downlink transmission. During downlink, MS also measures the channel and sends the Codec Mode Request (CMR) in the uplink transmission. In general, the codec mode control mechanism is located in the base station and only the base station can decide which mode should be used in the next speech frame. Two-bit Codec Mode Information (CMI) includes a maximum of 3 different modes. With our selection described before, we switch the modes between WB8.85, WB6.6 and NB5.15. We also would like to keep the original EFR (NB7.4) mode as a backup and compatibility mode. l l i e adaptation algorithm is based on channel measurements. In this work, we assume that the Cll value of the current channel has been measured correctly. With this assumption and with the result above, we found that the 13 dB and 18 dB SNR points are hvo obvious switching points. To ensure smooth codec mode transitions, an example of proposed switching decision thresholds of CMC and CMR is listed in Table 1.
For the simulations, we setup a testbed that simulates the end-to-end TDMA- 136 physical layer including N B N B AMR speech codecs, channel coding, interleaving, puncturing, the fading channel and n/4 DQPSK modulation. The channel we simulate here is the multipath Rayleigh-fading channel with additive white Gaussian noise (AWGN) and the testbed provides soft decision channel outputs with respect to a given vehicle speed. With this testbed, we simulate the NB5.15, NB5.9, NB7.4, WB6.6 and WB8.85 working in a TDMA 30 kHz channel with different channel SNRs. Listening test results are shown in Figure 3. To obtain these plots, randomized coded speech from the several bit rates and channel conditions were played through headphones for 25 untrained listeners. These listeners classified the speech as Excellent (S), Good (4), Fair (3). Poor (2), or Unacceptable (I). From Fig. 3, we can see that NB5.15 mode outperforms other modes when the channel SNR is below 12.5 dB. Behveen 12.5 dB and 17.5 dB SNR, WB6.6 has the best speech quality among all modes. WB8.85 takes over for better channels. To achieve the maximum system enhancement, we fmd the envelope of these curves as shown in Figure 3 and adapt codec modes at the cross points. It also shows that we can discard N35.9 and NB7.4 modes totally! By comparing the envelope and NB7.4 mode (EFR) performance, we can always get at least 1dB enhancement at the same speech quality. We also notice that our new system could achieve EFR quality of 15 dB at ahout 13.5 dB, which means that the TDMA-136 operational range is extended 1.5 dB with this approach.
-7 Figure 4. Proposed link adaptation scheme.
/
,
0
11
td
$7
18
I
Table 1 Proposed switching decision thresholds (Full Rate)
!I
Cham* OVlllt” (“,
Figure 3. Performance of AMR codec in TDMA channel.
r
5.2 Proposed Signaling and Link Adaptation Signaling between Base Station (BS) and Mobile Station ( M S ) for codec mode adaptation is in-hand signaling. The proposed system implementation scheme is shown in
556
MC’/MR’
)I THRMC/MRDn 1 THRlfC/MRUp 1 17dB
I
I
5.3
[SI "Adaptive Multi-Rate (AMR) speech transcoding", GSM 06.90, 1998.
Switching Between Narrowband and Wideband Speech
One important problem still left unsolved is that direct switching between narrowband speech and wideband speech may cause strong noise artifacts [Ill. The solution we proposed here is to switch between narrowband and wideband speech during unvoiced speech periods. However, this solution somewhat limits the switching speed and makes the adaptation more complex. Another novel solution that has been proposed based on bandwidth scalable spzech coding [ 1 I] can also be applied here.
[6] "AMR speech codec, wideband; Generaldescription" 3GTS26.171,2001.
6 Conclusions
[9] "AMR speech Codec; Error concealment of lost Games", 3G TS 26.091.
[7] "Mandatory Speech Codec speech processing functions AMR Wideband speech codec; Transcoding functions" 3G TS 26.190.2001. [SI "AMR speech Codec; Frame Structure" 3G TS 26.101.
A new TDMA-136 system enhancement approach is proposed based on the existing North American 136 30 kHz channel and NB/WB Adaptive Multi-Rate speech codecs. One advantage of this approach is that we change nothing in the time slot stTucture because the mode adaptation and mode information transmission all go through in-band bits. At both ends of BS and MS, the channel bandwidth is M y utilized by adaptively compromising between source and channel coding with puncturing. Another advantage is that the new TDMAAMR system outperforms the existing 136 system throughout a wide variety of channel conditions. Over I dB overall SNR enhancement is observed and the operational range is extended about 1.5 dB in our new system.
[IO] "Speech Codec speech processing functions; AMR Wideband Speech Codec; Error concealment of erroneow or lost frames" 3G TS 26.191 v5.0.0, 2001. I l l ] H. Dong and J. D. Gibson, "Bandwidth Scalable Speech Coding and Rate Switching" Proc. ZEEE ICASSP'OZ.2002.
References [I] N. R. Sollenberger, N. Seshadri, R. Cox, "The evolution of IS-I36 TDMA for third-generation wireless services" IEEE Personal Communiccrtions, Vol. 6. pp. 8-18, June 1999. [2] M. Austin, et al. "Service and system enhancements for TDMA digital cellular systenis" lEEE Pei-sonal Communications, Vol. 6, pp.20-33, June 1999. 131 T. Fingscheidf T. Hindelang, R. V. Cox and N. Seshadri, "Joint source-channel (De-) coding for mobile conununications" IEEE Transactions on Communications, Vol. 50, pp. 200-212, Feb 2002. 141 Digital Cellular Telecommunications System (Phase 2+), Channel Coding, GSM 05.03 version 8.5.0. Release 1999. June 2000.
557
