Improving quality of speech in VoIP using time-scale modification

Abstract

Voice over Internet Protocol (VoIP) is an emerging technology that enables the transport of voice over the Internet. The quality of delivered voice is a critical issue due to packet loss and end-to-end delay variations (commonly called jitter) introduced in the packet stream as it traverses the Internet. For non-real-time applications (such as Telnet, FTP, and email), these problems are not serious, as reliable delivery of data is ensured by TCP (Transmission Control Protocol). However, except in cases of high-bandwidth Local Area Network (LAN) connections, TCP cannot provide the loss and delay guarantees required by real-time applications. In VoIP, packet loss leads to the loss of portions of speech, resulting in poor voice quality. Packet loss at the receiver has two components: loss occurring in the network due to congestion at intermediate nodes, and loss at the receiver due to packets arriving later than their scheduled playout times. Since the emergence of packetized voice in the late 1970s, extensive work has been done to improve voice quality by reducing packet loss and mitigating its effects. In this thesis, we first propose a scheme to classify this work into two broad categories: packet-loss reduction schemes and packet-loss concealment schemes, with further sub-categories. Following this classification, we propose a fully receiver-side scheme to improve voice quality using a time-scale modification algorithm for speech, called the Global Local Search Time Scale Modification (GLS-TSM) algorithm. Without introducing additional buffering delays, the scheme reduces packet loss at the receiver by providing flexible arrival delay cut-offs for late-arriving packets. Additionally, it offers effective concealment of lost packets. The scheme operates by modifying the playout time-scale of successfully received packets already present in the receiver’s playout buffer. However, the original GLS-TSM algorithm was designed for large speech segments of 1600 samples (200 ms) and a fixed time-scale modification factor, ?. In VoIP, much smaller packet sizes are used - typically 10 to 40 ms. To modify the time-scale of these packets, parameters must be appropriately sized to avoid introducing extra packet holdup delays. Moreover, ? may vary on a per-packet basis depending on the packet arrival pattern. Therefore, the operational ranges of various GLS-TSM parameters were redefined to suit shorter signal segments and per-packet variability of ?. A novel scheme was also developed to calculate ? dynamically for each packet. Objective and subjective speech quality tests were conducted using a large number of input speech samples and widely varying network conditions to evaluate the performance of the proposed scheme. The speech samples were taken from a standard speech database, and the network was emulated using a tool called CloudFM. The quality of delivered speech was assessed using two standard objective speech quality measures. In addition to these, subjective quality tests were conducted to validate the predictions of the objective measures.