By Narciso Mera and Antoniette McDermott,
Lucent Technologies Microelectronics Group

With the worldwide acceptance of CDMA well under way, the focus is shifting from initial deployment to system enhancements and technology advancements. Service providers and equipment manufacturers are working in conjunction with the TIA to improve coverage, call quality and systems capacity. Latest evidence of this is the Enhanced Variable Rate Vocoder (EVRC), now being deployed in Korea.

The Korean service providers found themselves in a unique position in the deployment of CDMA. With 25% of the country’s population centered in Seoul, capacity planning had to be an underlying concern. Over a million subscribers in the first 15 months made the issue unavoidable. Competitive pressures between cellular and PCS service providers, meanwhile, meant that service quality enhancements also had to be considered.

At about the same time, the CDMA Development Group (CDG) and TIA were finalizing the specification of a new vocoder called Enhanced Variable Rate Vocoder (EVRC) or IS-127. The aim was to define a vocoder that provided the same or better quality than existing vocoders while improving system capacity.

The original CDMA voice coder, IS-96A, was an 8kbit/s Code-Excited Linear Predictive (CELP) coder designed for use in the digital cellular band. The desire for improved voice quality spurred the TIA to begin working on a 13kbit/s CELP vocoder to provide higher quality voice transmissions. The 13kbit/s CDMA vocoder takes advantage of the higher data rate to improve speech quality. It has shown mean opinion scores (MOS) consistent with toll quality voice, the benchmark for comparison. Unfortunately, capacity and cell site range are affected by the higher data rate vocoder.

The latest CDMA vocoder, EVRC, takes advantage of advances in signal processing hardware and software techniques to provide 13kbit/s voice quality at 8kbit/s data rates, thereby maximizing both quality and systems capacity.

MOS data has shown that 8kbit/s EVRC compares favorably to 16kbit/s LD-CELP and ADPCM, the industry standards for comparison. More importantly, these tests show that EVRC maintains superior quality over 13kbit/s CELP as frame error rates rise.

DESIGN CONSIDERATIONS
One of the important and unique aspects of CDMA wireless systems is that while there are limits to the number of phone calls that can be handled by a given carrier at a time, this capacity limit is not a fixed number. Rather, the fundamental physics of CDMA are such that cell coverage depends on the way the system is designed and implemented. Systems capacity, voice quality and coverage are all interdependent, enabling an operator to trade any one off against the other two.

To maximize the number of simultaneous calls that can be handled at any given time in the allocated frequency spectrum, digital wireless systems utilize speech compression between the mobile and base station. This is especially important in CDMA, where multiple conversations are spread across the same spectrum and differentiated by pseudo random codes. In essence, each signal or voice call appears as noise in the system to the other calls. The higher the compression ratio (or the fewer number of bits), the lower the noise contribution of each call, which thereby improves system capacity.
The challenge is to represent each call with the fewest bits without sacrificing voice quality and system cost. Voice quality is determined primarily by the speech coding technology used.

EVRC FUNDAMENTALS
The core of the standard EVRC algorithm is a 1994 algorithm called Relaxed Code-Excited Linear Predictive coding (RCELP). It was developed by Bell Labs, Lucent’s research and development arm. RCELP, a generalization of the CELP speech coding algorithm, is particularly well suited for variable rate operation and robustness in the CDMA environment. CELP uses 20 millisecond (ms) speech frames for coding and decoding. In each 20ms time interval, the encoder processes 160 samples of speech.
With variable rate coders, the encoder examines the contents of each speech frame and determines the necessary coding rate. Depending on the voice waveform (volume, pitch, rate and so on), the coder represents speech at one of three bit rates: 8kbit/s, 4kbit/s, or 1kbit/s. As a result, the average bit rate is less than 8kbit/s. This differs from a half-rate or multi-rate coder, where the bit rate is determined once for each call. In addition, when no voice is detected, the vocoder drops its encoding rate and the effective bit rate goes to 1kbit/s, further reducing the interference energy produced.
CELP coders use three sets of bits to represent speech: linear predictor filter coefficients; pitch parameters; and excitation waveform. For each 20ms frame, the CELP algorithm examines the data and generates 10 linear prediction coding filter coefficients represented with differing bit precision, depending on the variable rate. With EVRC, the coefficients are represented by a vector, which is a set of the most likely usable coefficients. Increased or decreased bit precision is applied as necessary.

CELP speech coders also perform long-term pitch analysis that generates a 7-bit pitch period and a 3-bit pitch gain. The analysis here is based on a mathematical model of the human vocal track. At the different rates, the coder performs this pitch analysis on either four 5ms subframes, two 10 ms subframes or one 20 ms frame. The result is a variable number of bits per frame representing the pitch information. EVRC makes two pitch measurements and uses warped pitch delays, removing the requirement for a large number of bits and increased computations for fractional pitch delays.

IMPROVED QUALITY
The excitation waveform for every frame or subframe is selected from a codebook consisting of a large number of candidate waveform vectors. The codebook vector chosen to excite the speech coder filters minimizes the weighted error between the original and the synthesized speech. The techniques employed by EVRC for reducing the number of bits required for linear predictor coefficients and pitch synthesis enable the use of a more extensive algebraic codebook to generate the excitation. As a result, EVRC has higher voice quality.

Unlike conventional CELP encoders, EVRC does not attempt to match the original speech signal exactly. Instead, EVRC matches a time-warped version of the residual that conforms to a simplified pitch contour. The contour is obtained by estimating the pitch delay in each frame and linearly interpolating the pitch from frame to frame. While this adds to the computational complexity, the result is higher voice quality per bit transmitted.

The simplified pitch representation also leaves more bits available in each packet for the stochastic excitation and channel impairment protection than would be possible if a traditional fractional pitch approach were used. The result is enhanced error performance without degraded speech quality at the small cost of added processing requirements.

EVRC also enhances call quality by suppressing background noise. The IS-127 standard recommends a noise suppressor algorithm, but allows system designers to define their own. This is an important factor in choosing a processing platform, making programmable DSPs, such as Lucent Technologies’ DSP offerings, a desirable choice.

SERVICE PROVIDER DILEMMA
By early 1997, it started to become clear to service providers that EVRC was an attractive choice for addressing their capacity and voice quality requirements. The challenge then became how to make EVRC a reality, given that existing equipment, handsets and infrastructure did not support the technology. Further complicating the issue, service providers had a significant investment in existing infrastructure and handset manufacturers were just beginning the ramp-up of new designs.

The Korean market was the first to take the initiative in addressing this dilemma. In an effort to realize EVRC quickly and minimize the cost of the development effort, several Korean manufacturers worked closely with Lucent Technologies to define and implement a strategy.

Lucent Technologies provided a complete EVRC solution for CDMA manufacturers based on its DSP1627 chip. To quickly enable deployment, the implementation of EVRC leveraged the existing equipment investments as much as possible. Separate solutions were required for the handset and infrastructure and each presented different challenges Figure 1 illustrates a typical CDMA wireless network and highlights where EVRC fits into the overall architecture.

For the CDMA handsets, Lucent’s EVRC chip can be used as a standalone DSP or as a co-processor to the existing architecture as illustrated in Figure 1. The integrated solution not only includes the device, but also all the necessary software to enable EVRC. Additionally, Lucent’s implementation is programmable, offering handset manufacturers ample opportunity to tune or differentiate their solutions to address particular market needs.

Lucent’s implementation of EVRC in the infrastructure was a software upgrade to the existing speech coder platform, rather than a complete redesign and switch out. This was the most desirable solution for infrastructure due to the fact that replacement of deployed equipment was cost prohibitive. The task was non-trivial since the existing platform had been designed to the performance requirements of a less complex speech coder.

This massive effort was a major achievement: concept, specification, development, integration, field trials and production ramp were completed in less than six months. This could not have been done without a coordinated effort between the service provider, the equipment manufacturers and Lucent Technologies.

FUTURE OUTLOOK
EVRC is the first of many developments in the industry’s push to close the gap between wireless and wireline communications. Other developments include quality, cost, services, ease of use, and availability.

Today, the Korean market is driving the deployment of EVRC. As CDMA technology deploys in other regions of the world, service providers may face the same dilemma — competitive pressures and systems capacity. Given the demonstrated benefits of EVRC, they may choose to address the challenge in the same manner. Although, with the rapid pace of technology advancements in both speech coding algorithms and signal processing hardware, it’s likely that more alternatives will become available.

On the signal processing front, the next generation of DSP architectures is now available. Lucent recently introduced its DSP16000 family, for instance, which is designed to specifically address the future needs of the wireless communications market. It offers more than twice the performance of current DSPs with reduced power consumption and costs. This level of processing capability will be crucial in the development of even better speech coders, with advanced features such as voice recognition and high speed data services.