Measure Audio Quality (DNSMOS)

@microsoft1 VersionEstimate

Scores each AudioFrame with the DNSMOS network and emits four MOS-style quality numbers (speech, background, overall, and a P.808 estimate), shaped either as a named record or a [Analytics.Metric] list depending on what the downstream stage expects.

How it fits

Typical backends

Audio is scored and the record-form result is exported over HTTP for live monitoring or quality dashboards.

Quality is checked before speech recognition so low-grade audio can be gated out ahead of transcribe_audio_faster_whisper.

Caveats

AccuracyThe numbers are model-predicted MOS approximations, not subjective listener scores, so they are best used to compare relative quality rather than as absolute values.
I/O contractThe output shape is fixed when the graph is built from the downstream connection: a record for direct gating or a metric list for aggregators; an unsupported downstream type fails at startup.
FallbackAn empty input frame returns all four scores as zero with no error, so a gate that keys on a non-zero overall score will quietly drop empty frames.
CompatibilityFrames are resampled internally to the model's 16 kHz and mixed to mono, and requesting GPU through device quietly runs on CPU when no GPU is present.

Versions

02716a9clatestdefault
Automated release
6/25/2026