Gradual Decoding Refresh

Description

Gradual Decoding Refresh (GDR) is a mechanism implemented within video and speech codecs to refresh the decoder's state incrementally rather than instantaneously. In traditional predictive video coding (e.g., H.264/AVC, H.265/HEVC), a decoder relies on reference pictures stored in a Decoded Picture Buffer (DPB). If packets are lost or if a new viewer joins a stream (a "random access" point), the decoder's reference buffers may become corrupted or mismatched with the encoder's, leading to persistent visual artifacts. The typical recovery method is an Instantaneous Decoding Refresh (IDR) frame, which is a large intra-coded frame that completely resets the decoder's state, but at a high bitrate cost and potential latency.

GDR works by encoding a sequence of pictures in a specific way that allows the decoder to gradually construct a clean reference picture. Instead of a single large IDR frame, the encoder spreads the refresh information across multiple consecutive frames. One common technique involves partitioning a frame into slices or regions and encoding each region as an intra-coded block (I-block) over several frames in a cyclic manner. For example, over a cycle of N frames, each frame might refresh 1/Nth of the total picture area with intra coding, while the rest of the picture is predictively coded (using P or B frames). As the decoder receives these frames, it progressively replaces corrupted or old parts of its reference picture with the newly intra-coded blocks. After receiving the complete cycle of N frames, the decoder's reference picture is fully refreshed with intra-coded data, achieving a state equivalent to having decoded an IDR frame, but without the large instantaneous bitrate spike.

Within 3GPP, GDR is specified for video services and is a key feature of the Enhanced Voice Services (EVS) codec, as documented in TS 26.906 (EVS Codec) and TS 26.926 (EVS Codec test sequences). For EVS, which handles super-wideband and full-band speech, GDR-like mechanisms are applied to the parameter domains of the speech codec to allow graceful recovery from packet loss and smooth switching between different operating modes (e.g., between voice activity detection (VAD) and continuous transmission). The codec can gradually update its internal state (e.g., LPC filter coefficients) rather than forcing an abrupt reset, which maintains higher perceptual speech quality during network impairments or handovers. In video services, GDR enables more efficient adaptive bitrate streaming and improved error resilience in mobile environments where packet loss and delay are common, directly contributing to a better Quality of Experience (QoE) for end-users.

Purpose & Motivation

GDR was developed to solve the fundamental trade-off between random access/error recovery efficiency and coding efficiency in predictive media coding. The classic solution, the IDR frame, provides a clean decoder reset point but is extremely inefficient in terms of bitrate. In low-latency real-time communication (RTC) scenarios, such as mobile video calls or voice-over-LTE (VoLTE), frequently sending large IDR frames to combat packet loss would consume excessive bandwidth and introduce noticeable quality dips or latency. Conversely, not sending IDR frames often enough leads to prolonged picture corruption or desynchronization between encoder and decoder after packet loss. GDR provides a middle path, allowing the decoder to recover gradually without a large instantaneous bitrate penalty.

The historical context stems from the evolution of video compression for streaming and broadcasting, where channel switching times and error resilience are critical. Earlier standards like MPEG-2 used periodic intra-frames (I-frames), which are similar to IDR frames. GDR mechanisms were refined in later codecs like H.264 and became more prominent with the rise of Internet and wireless video delivery, where bandwidth is variable and losses are frequent. For 3GPP, the adoption of GDR in specifications like those for EVS was motivated by the need for robust, high-quality voice and video services over LTE and 5G NR networks. These networks, while high-capacity, still experience radio link variations, handovers, and temporary congestion.

Specifically, for the EVS codec introduced in 3GPP Release 12, GDR techniques address the challenge of mode switching and packet loss recovery in a bandwidth-efficient manner. EVS supports a wide range of bitrates and operation modes (e.g., speech, music, mixed content). Switching between these modes, or recovering from a burst of lost packets, requires re-synchronizing the decoder's internal state. A full reset would cause an audible glitch. GDR allows this synchronization to happen smoothly over several frames, preserving voice naturalness. Thus, GDR solves the problem of maintaining continuous, high-quality media delivery in dynamic and potentially lossy mobile network conditions, enabling services like VoLTE and ViLTE to meet user expectations for clarity and reliability.

Key Features

• Enables decoder state refresh without a full instantaneous decoding refresh (IDR) frame
• Spreads intra-coded information across multiple consecutive video frames or speech packets
• Reduces peak bitrate requirements compared to periodic IDR frames, improving bandwidth efficiency
• Provides seamless random access and error recovery, minimizing visual or audible artifacts
• Integral to the 3GPP Enhanced Voice Services (EVS) codec for robust speech transmission
• Enhances error resilience and Quality of Experience (QoE) in loss-prone mobile networks

Evolution Across Releases

Defining Specifications