Year: 2024
Type: dissertation
Abstract: Significant advancements in audio transport, encoding, and processing power in embedded systems constitute the foundation of today’s speakers. Transitioning the digitalized audio data transmission to Audio over IP (AoIP) has opened the potential to transmit audio with the same network flexibility as other services using packet-based networks, such as voice over IP or video over IP. This thesis primarily focuses on evaluating different masking techniques for transient network errors in embedded environments, such as an IP speaker using a simulated audio pipeline. The key indicators of success are the algorithms’ latency, delay, perceived audio quality, and complexity. Both sender-based and receiver-based, as well as combinations of these methods, have been evaluated. Two techniques were sender-based (Redundant Transmission (RT) and Parity Packet Forward Error Correction (PPFEC)). Five were receiver-based (Silence Insertion (SI), frequency-dependent White Noise Insertion (WNI), Packet Repetition (PR), Waveform Substitution based on Pattern-matching (WSP), and Waveform Similarity Overlap and Add (WSOLA)). The sender-receiver-based techniques were the 10 resulting combinations of sender-based and receiver-based techniques. Considering all the key indicators, the findings were that the best-performing algorithms for receiver-based Packet Loss Concealment (PLC), sender-based Forward Error Correction (FEC), and the combination sender-receiver-based were PR, PPFEC, and PPFEC with PR. PR had a mean opinion score (MOS) of 57.08 across all evaluated tracks at a drop percentage of 10% with a segmented crosscorrelation score roughly at 0.5, which, compared to SI, only had a segmented similarity score in the range of 0.1 to 0.3 for the evaluated excerpts. Further, the execution time ratios between SI and PR were almost equal. Introducing FEC into the masking techniques further improved the results since the reconstruction made by FEC techniques perfectly reconstructs the lost segment if enough redundant data was transmitted correctly. However, implementing FEC using RT or PPFEC results in increased end-to-end latency due to the increased amount of redundant information transmitted, with RT causing a greater increase than PPFEC. The conclusion was that when latency does not have to be ultra-low, the combination of PPFEC and PR will do a great job of masking occasional transient errors in low-resource AoIP-embedded systems. In the case of ultra-low requirements, using only PR is suggested since PR has low complexity, introduces no new latency, and does not further strain the network.
Author: Midlöv, Alexander
Institution: Lund University
Language: English
Topic: Speech and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7001968293