Quality monitoring & MQA
The encoder is the only component in the chain that has both the source and the encode at the same time. Norsk tags every frame with quality metadata as it encodes, aggregates it into per-segment scores, and carries it downstream via CMSD standard headers. Every packager, CDN, and monitoring tool reads the same quality signal — without decoding the media, without DRM keys, without rebuilding what was already calculated. Norsk is among the authors and earliest implementers of the CMSD-MQA specification.
How it works
Quality is measured once, at the point it is cheapest to calculate, and carried as structured metadata all the way to the player. Every component reads the same signal. No component rebuilds it.
Full-reference metrics like VMAF, PSNR, and SSIM need both the source signal and the encoded output, compared frame by frame. The encoder is the only place in the chain where both exist simultaneously. Everywhere downstream, the source is gone — so any system that wants a quality reading has to rebuild it by decoding the stream, re-running the analysis, and, for protected content, holding DRM keys. Norsk eliminates that cost by tagging every frame with quality metadata at the point of encode. Fidelity scores are embedded in SEI messages inside the H.264 stream. Transport-level checks — continuity counter errors, dropped frames, decoder faults — are tagged at the same level. The information is created once, at the moment it is free to create.
The packager aggregates frame-level scores into per-segment quality scores and writes them into CMSD-Static response headers. The SVTA2128 specification mandates two composite numbers: an overall MQA video score and an overall MQA audio score, each 0 to 100, higher is better. Linear aggregation is standard — so a packager from one vendor reads frame data from another vendor's encoder and produces the correct segment score without a proprietary integration. The scores ride in standard HTTP headers and survive every CDN and cache in the delivery chain without opening the media. A segment can post a high PSNR from a fidelity perspective and still score 54 overall if transport errors are counted — one number that reflects the whole picture, with the constituent detail available behind it.
With MQA scores flowing in real time, every component in the chain can make decisions on actual picture quality rather than bitrate assumptions. An origin shield pulls each segment from the higher-scoring source when a redundant feed degrades. A packager routes away from a rain-fade contribution before the viewer sees the pixelation. A monitoring dashboard surfaces the quality drop on the first affected segment, with the cause already identified. A player records what the viewer actually received — an audit trail that closes the loop from encoder to screen. Hysteresis is applied at the switcher so the feed does not change lanes on a single low-scoring segment; several good segments are required before reverting to the preferred source. Override is always available.
Use cases
The operations that benefit most are those where the gap between 'the encoder knows it' and 'the operator knows it' is where faults propagate, costs accumulate, or compliance questions go unanswered.
A broadcaster takes contribution feeds from two satellite uplinks into separate encoding regions. One feed periodically suffers rain fade — the picture pixelates. But VMAF and PSNR scores can actually rise when rain fade hits: pixelated content has less information, so the same bandwidth encodes less detail more faithfully, and the fidelity metric improves while the picture falls apart. A system relying solely on fidelity metrics routes toward the broken feed with confidence.
Norsk combines fidelity scores with transport-level filters — sequence errors, dropped frames, decoder faults. The composite MQA video score drops immediately when rain fade degrades the signal, regardless of what VMAF reports in isolation. The origin shield switches to the clean region automatically, on the first affected segment, without operator intervention. No probe at the origin, no DRM key distribution, no delay waiting for an alert to reach a human.
Talk to us about your contribution workflow →A large-scale streaming operation monitors quality with a probe farm distributed across the delivery chain. For premium content, every probe that wants a quality reading needs DRM keys — a cost centre and a security surface. The probes are expensive to operate at scale, and they are mostly recalculating information the encoders already produced and then discarded.
Quality is computed once, at the encoder, and carried as CMSD headers through the chain. Downstream systems — packagers, CDNs, monitoring dashboards — make routing and alerting decisions by reading the header, not by decoding or decrypting media. The probe farm reduces to dashboards consuming structured metadata. DRM keys stay where they belong: with playback systems. The cost centre and the security exposure leave at the same time.
Talk to us about your monitoring setup →A streaming service sells HD subscriptions and inserts ads from third-party sources into live streams. It cannot confirm, segment by segment, that subscribers received HD quality rather than a degraded stream — or that inserted ads matched the perceptual quality of the programme around them. Separately, with modern codecs on simple content, adjacent rungs of the bitrate ladder are often perceptually identical, but without quality data in the chain, delivery decisions stay conservative.
MQA scores in the manifest give the service a segment-by-segment audit trail of what quality every viewer actually received. The HD subscriber data is objective, not inferred from bitrate. For ad-quality verification, each ad segment carries its own score. On delivery efficiency: when the quality difference between adjacent bitrate rungs is negligible, the lower one is served — the saving lands directly on the CDN bill, verified rather than estimated.
Talk to us about quality audit and delivery →Capabilities
SVTA2128 is published. AWS Elemental MediaPackage already emits CMSD quality scores. Norsk demonstrated live interoperability at NAB 2026 with Akamai, Touchstream, and G&L Systemhaus.
Full-reference metrics require both the original source and the encoded output. The encoder has both. Downstream, the source is gone. Norsk calculates VMAF, PSNR, and SSIM at encode time and embeds per-frame scores in SEI messages inside the H.264 stream — available at frame, GOP, or segment granularity for any downstream system that needs them.
Rain-faded content can post high VMAF while the picture falls apart: pixelation reduces information, so the same bandwidth encodes less detail more faithfully, and the fidelity metric improves. Norsk combines fidelity scores with transport-level filters — continuity counter errors, dropped frames, decoder faults, audio out of range — into the composite MQA video and audio scores. A segment can post PSNR 47 and SSIM 100 and still score 54 once transport errors count. One number reflects the whole picture.
Per-segment MQA scores ride in CMSD-Static response headers. The data is payload-agnostic: it works across any CDN, any common-encryption scheme, live and on-demand, as long as the media is adaptive and segmented. No proprietary protocol, no specialist client library. Existing caching infrastructure, CDNs, and monitoring tools read the same header without modification.
Downstream systems read quality scores from CMSD headers — they do not decode or decrypt the media to produce them. For premium and DRM-protected content, monitoring and routing decisions happen entirely on metadata. DRM keys stay with playback systems. The key surface and the probe farm cost centre both shrink.
SVTA2128 defines how frame scores aggregate into segment scores with linear averaging, so a packager from one vendor reads frame data from another vendor's encoder and produces the correct result. Norsk demonstrated this live at NAB 2026 with Akamai, Touchstream, and G&L Systemhaus — automated quality-based switching across four independent products, all reading the same CMSD headers.
Quality-based switching between redundant paths is only safe when both paths cut segments at the same frame. Repeating or skipping even two seconds during a switch is worse for the viewer than the glitch the switch was meant to avoid. Norsk injects SEI time codes at the source so encoders in different regions make identical segmentation decisions without coordinating — demonstrated at 24-hour scale with frame-accurate alignment between two independently-started encoders.
Norsk is among the authors and earliest implementers of the CMSD-MQA specification. If you are evaluating quality monitoring, building a resilient contribution workflow, or looking to reduce probe infrastructure, get in touch.
Talk to us