Description
The DSR Optimised Codec is a speech codec specifically designed for Distributed Speech Recognition (DSR) applications. Unlike conventional speech codecs such as AMR, which are optimised for human listening and aim to reproduce the original waveform as faithfully as possible, the DSR Optimised Codec is optimised for machine recognition. It takes speech input, extracts acoustic features (such as mel-frequency cepstral coefficients, MFCCs), and encodes them into a bitstream that can be directly consumed by a speech recognition engine on the network side. This eliminates the need to reconstruct the waveform at the server, reducing computational complexity and latency.
The codec resides in the User Equipment (UE) and operates as part of the DSR framework. In a typical DSR architecture, the UE captures the user's speech, processes it through the DSR Optimised Codec, and transmits the resulting feature stream over the 3GPP network to a remote speech recognition server. The server-side speech engine then performs recognition directly on the feature stream. This split architecture is beneficial because it offloads the computationally intensive feature extraction from the server to the UE, and it reduces the required uplink bit rate compared to transmitting compressed speech and then extracting features at the server.
The DSR Optimised Codec is defined in 3GPP specifications 22.243 (Speech recognition framework for automated voice services) and 22.977 (Study on speech recognition framework). These specifications describe the requirements for DSR, including the codec's performance, bit rate, and robustness to transmission errors. The codec is designed to be robust against packet loss and bit errors, which are common in mobile networks. It may include error protection mechanisms such as forward error correction (FEC) and frame interleaving. The codec also supports discontinuous transmission (DTX) and comfort noise generation to save resources during silence periods.
From a network perspective, the DSR Optimised Codec is treated as a media type in the 3GPP QoS framework. The network must provide appropriate QoS to ensure that the feature stream is delivered with low latency and low loss. This may involve using a dedicated bearer with guaranteed bit rate (GBR) and a suitable QCI/5QI. The codec's bit rate is typically low (in the range of 2-5 kbps) compared to conventional speech codecs, which makes it suitable for bandwidth-constrained scenarios. However, the codec's performance is sensitive to latency, as the speech recognition engine may require real-time or near-real-time processing.
The DSR Optimised Codec is an example of how 3GPP networks can support innovative services beyond voice telephony. It enables applications such as voice dialling, voice search, and voice-controlled assistants to operate efficiently over mobile networks. By standardising the codec, 3GPP ensures interoperability between different UE vendors and speech recognition service providers, which is essential for widespread adoption. The codec has been part of 3GPP specifications since Release 8 and has been carried forward through subsequent releases, reflecting its continued relevance in the era of smart assistants and AI-driven voice services.
Purpose & Motivation
The DSR Optimised Codec was created to address the inefficiencies of using conventional speech codecs for speech recognition applications. In a conventional approach, the UE would compress the speech using a codec like AMR, transmit it to a server, and the server would then decompress the speech and extract acoustic features for recognition. This approach has several drawbacks: it requires the server to perform both decompression and feature extraction, which increases computational load and latency; it consumes more uplink bandwidth than necessary; and it may degrade recognition accuracy due to compression artifacts.
The DSR Optimised Codec solves these problems by moving the feature extraction to the UE and encoding the features directly. This reduces the bit rate, lowers latency, and improves recognition accuracy because the features are extracted from the original speech rather than from a compressed version. It also simplifies the server architecture, as the server can directly interface with the feature stream. The codec was motivated by the growing demand for automated voice services, such as voice dialling and voice search, which require efficient and accurate speech recognition over mobile networks.
The standardisation of the codec in 3GPP ensures that different UEs and servers can interoperate, which is crucial for the success of voice-based services. It also provides a basis for performance testing and certification. The codec's design considers the characteristics of mobile networks, such as variable bandwidth and error-prone transmission, making it suitable for real-world deployment.
Release Timeline
Evolution Across Releases
An overview written from the defining specifications, not a verified change history. For the exact text, open the specifications listed below.
Initial definition introduced in 3GPP specifications 22.243 and 22.977: "The module in UE which takes speech input, extracts acoustic features and encodes them with a scheme optimised for speech recognition. This module is similar to the conventional codec (e.g. AMR). On the server-side, the uplink encoded stream can be directly consumed by speech engines without having to be converted to a waveform." This established the DSR Optimised Codec as a key enabler for distributed speech recognition over 3GPP networks.
Defining Specifications
3GPP specifications that define or reference DSR Optimised Codec, with the latest known release. Sourced from the 3GPP document catalog — see methodology.
| Specification | Title | Release |
|---|---|---|
| TS 22.243 vj00 | Speech Recognition Framework for Automated Voice Services | Rel-19 read |
| TR 22.977 vj00 | Speech Enabled Services and Multimodal Framework | Rel-19 read |