With a rapidly growing number of 5G networks deployments, and the GSA now reporting that over 1,100 5G devices have been announced globally, the launch of new voice services over 5G New Radio (VoNR) is becoming the top priority for operators. But responding to this demand creates a challenge: maintaining the voice service quality and growing voice revenue through VoLTE expansion while minimizing CAPEX and OPEX.
In addition, the continued strong competition of carrier-grade voice services (VoLTE, VoNR) with OTT voice applications is predicted to grow exponentially with 5G network deployments, presenting a continuously persistent threat to MNOs’ voice service revenue. The powerful MNOs’ counter to this threat is the expansion of the carriers’ VoLTE/VoNR services, while continuously monitoring and benchmarking against OTT voice applications to maintain their voice services at high QoE levels.
In this blog post, we:
- Reintroduce the concept of sQLEAR, now that it has been recognised as an ITU-T standard;
- Discuss its differentiators, performance, and field operability; and
- Share the main benefits for users, operators and regulators alike.
Infovista’s ITU standardization effort
Having envisioned this evolution – or trend – of 4G/5G voice services, in 2018 Infovista introduced its sQLEAR (speech Quality by machine LEARning) machine learning-based algorithm for voice QoE testing of mobile IMS voice services (e.g. VoLTE, VoNR), and OTT voice.
Furthermore, strongly convinced of the importance of re-thinking voice quality testing with ML-based solutions and network-centric evaluation of the user voice QoE, Infovista developed within ITU a framework for voice service quality monitoring and troubleshooting for intrusive parametric voice QoE prediction using ML. This work has been finalized with an ITU recommendation: ITU-T P.565, first published in January 2020.
As a long-time pioneer in the delivery of voice QoE assurance solutions, Infovista furthered its activity working closely with ITU experts and independent parties to test and validate our solution, sQLEAR, as the first outcome of ITU-T P.565 framework.
During this process, extensive validation was performed by the ITU on a large variety of databases collected from real-life mobile IMS networks (VoLTE, VoNR), and OTT/WhatsApp, as the most used OTT app.
sQLEAR proved to meet the ITU-T performance requirements for all the validation databases: the performance requirements, as defined in ITU-T P.565, are of 95% correlation, with a root mean square error of 0.35MOS and a mean absolute error of 0.3MOS against the user-perceived quality. In addition, 95% of all tested voice samples should show absolute errors lower than 1MOS.
Consequently, as we announced last week, sQLEAR became a new ITU recommendation: ITU-T P.565.1, the first machine learning standard for mobile IMS voice (VoLTE, VoNR) and OTT voice QoE testing.
sQLEAR uniquely combines the time characteristic of the input reference voice sample to be sent over the network with the network/client/codec parameters. Therefore, sQLEAR represents a unique QoE model, which is intrusive and parametric. The intrusiveness reflected by the active testing with the reference voice sample empowers sQLEAR with significantly increased accuracy versus a solution based only on parameters.
sQLEAR is based the following main factors:
- Transport protocol information, such as jitter and packet loss;
- Codec information, which includes bit rate, bandwidth and modes (e.g. channel-aware and interoperability modes for VoLTE/VoNR EVS codec case);
- Client information (it has a standardized jitter buffer error concealment scheme); and
- Input voice reference to be transmitted over the network.
The sQLEAR algorithm uses deep-packet inspection (DPI) to obtain this relevant information, which means that the impact of the network on voice QoE can be uniquely determined without the need to record actual voice content. This ensures that sQLEAR field measurements are less sensitive to any unwanted test artifacts.
In addition, using the time characteristics of the reference voice sample enables sQLEAR to accurately identify the importance of individual sections of the bitstream in regard to voice quality as perceived by users. This also offers the advantage of considering the real voice signal after the voice client jitter buffer, which ensures that sQLEAR can predict the quality of experience of the voice service delivered over a network working with a standardized device-based voice client.
Furthermore, using bitstream data allows sQLEAR to be minimally sensitive to language and speech frequency dependencies.
Uniquely, sQLEAR uses state-of-the-art machine learning to build a model that describes the voice quality perceived by users based on all of these information resources. The ML techniques offer a significant advantage versus the multi-dimensional optimization techniques required for the estimation of coefficients of multi-variable non-linear functions, which are used in traditional QoE modelling.
The advantage comes from the fact that ML techniques are best suited to describe the increased complexity of the inter-dependencies between all network/codec/client parameters, which comes with mobile IMS voice and especially 5G New Radio, as well as the significance of these interdependencies on impacting the voice quality of experience.
Read more technical details on how sQLEAR works in our white paper, Learn about sQLEAR in motion.
Using sQLEAR in field measurements
As an active voice test, sQLEAR uses the same test setup as defined by ITU-T/ETSI, but without the need for the recorded degraded voice sample, as explained above. In addition, during run-time, sQLEAR runs a pre-processing phase during which the reference voice is synchronized with the corresponding IP/RTP bitstream to ensure that it works on the exact bitstream of the reference voice sent through the network. If desired, recorded speech can be saved for further offline analysis.
Furthermore, during the pre-processing phase, several other tasks required by ML techniques are applied. The output of the pre-processing phase is given to the pre-trained sQLEAR algorithm, which outputs the predicted score of the voice quality. A set of parameters meaningful to root cause analysis of the voice quality can be also reported.
By running sQLEAR in their field measurement tools, operators can accurately identify voice quality trends as perceived by their users. For the convenience of backwards compatibility, these trends can be compared with the ones that come from voice quality measurements performed with traditional perceptual QoE models. In addition, since sQLEAR uniquely provides the impact on the voice quality of networks working with devices using standardized voice clients, well-designed apple to apple comparisons of individual sQLEAR scores’ (per individual voice samples) distribution with scores’ distribution determined using traditional perceptual voice QoE metrics can help operators identify the best performing devices working with their network.
Read more on how sQLEAR positions itself versus other available voice QoE metrics developed with ITU-T in our white paper, A new approach for testing voice quality: sQLEAR Q&A.
The benefits of sQLEAR for operators and regulators alike
Uniquely, sQLEAR bundles ML techniques, network parameters, standardized voice codec and client information, and uses machine learning to provide mobile operators with the network-centric, device-agnostic, audio path-independent, real-time view of the true voice quality being delivered through their 4G and 5G networks. This significantly reduces both cost and time to market of new 5G voice services, while cost-efficiently maintaining high quality standards for existing VoLTE services.
Free of devices’ audio path impact, sQLEAR empowers operators with cost-effective monitoring, optimization and troubleshooting of their 4G and/or 5G network, without the need to individually test all commercial devices. This enables operators to save time and money, both by optimizing their networks for all, rather than for specific, devices and by being able to quickly identify any network-based issues without interference from device characteristics, which could mislead root-cause analysis of the voice quality.
In addition, as a device-agnostic voice quality predictor, sQLEAR provides operators and regulators alike with an accurate and fair benchmarking tool. Furthermore, without the need to work with the voice signal, sQLEAR is more versatile for various languages, and consequently suitable to use in different world regions.
sQLEAR exploits ML capabilities to most suitably describe the impact on voice quality of the increasingly complex network, voice codec and client interdependencies that are inherent in the mobile IMS voice networks (VoLTE, VoNR). Therefore, users of sQLEAR benefit from a trusted and reliable predictor of 4G/5G voice quality of experience. Furthermore, sQLEAR accuracy performance and reliability is enforced by its ITU-T approval as a standard.
Last, but not least, field sQLEAR deployments are intuitive, easy to use and, to a large extent, what users are accustomed to running when it comes to voice quality testing, but with two bonuses: no need to access the audio path and record the voice signal; and with the possibility of automatically getting the root-causes of the voice quality along with the sQLEAR score.
Infovista’s sQLEAR is the first industry solution and ITU-T standard to utilize a combination of machine learning, network/client/codec information, and reference audio for the assessment of transmission network impact on voice quality for mobile all IP voice services. Therefore, sQLEAR is the first intrusive parametric voice QoE predictor in the industry, which can be accurately and reliably used for high-definition (super-wide and full band) voice QoE testing across 4G/ LTE and 5G networks, both carrier (a.k.a VoLTE, VoNR) and OTT (e.g. WhatsApp).
Watch this space for further sQLEAR-themed blog posts.