Communication scored like physics, not vibes.
Most tools ask an AI to "rate communication skills" and get back an adjective. We refused to ship an adjective.
Our engine measures seven delivery signals straight off the clock and the waveform, then blends them with what was actually said. Every number is countable. Every claim carries a playable clip as evidence.
Analysis schema sharedpro.communication.delivery.v1
Delivery, decomposed into countable signals.
Each metric is scored against an optimal band. Speaking too fast loses points just as speaking too slowly does, while being too animated loses points just as monotony does. The weights shown represent each meter's share of the delivery composite.
Articulation rate
Articulation rate uses content words. The report supplies its exact unit, score availability, evidence, and caveat.
Mid-clause halts
Pauses are classified by location: stopping between clauses is treated as thinking and is never penalized. Only halts inside a thought are counted; each halt of ≥2s gets an evidence clip.
Filled pauses
The ums and uhs, word-level timestamped. Soft fillers ("like", "basically") only count when they exceed 3/min. Counts are treated as a floor, and the report says so.
Self-repairs & restarts
Immediate word repetitions and phrase restarts, such as "I'd... I'd approach," are detected from the token stream per 100 words.
Were the words understood?
Accent-safe by construction: it measures whether speech was understood through ASR word-recognition probability, never how it was pronounced. A microphone-quality caveat accompanies the number when relevant.
Pitch variety
F0 spread is measured in semitones relative to the speaker's own median, never in absolute Hz, so a deep voice and a high voice sit on the same scale. Scores taper when speech is either too flat or too erratic.
Response latency
Median seconds from the end of a question to the onset of speech. Suspiciously uniform 6–9s latencies don't affect this score; they feed the assisted-answer check instead.
Time to first point
Measures how long it takes for a substantive point to appear, as located by the content model in the transcript and read from word timestamps. Reported for each answer alongside duration and word count.
Withheld beats guessed
Thin audio reports values but withholds scores. Cohort percentiles are not computed from fewer than 30 samples. Coverage accompanies every composite; a blank cell is information, while a guessed one is a liability.
Bands shown are pre-calibration defaults · fitted against a human-rated anchor set before any evaluative use · every report records its scoring version
Run the meters on your own 30 seconds.
Answer any interview question out loud, and the same pipeline you just read about scores it. There is a quick human check first so bots don't hog the GPU.
- Pass the human check, which is invisible for most people.
- Hit play to hear a 30-second communication sample.
- Get your seven meters: values, band scores, and what was withheld. You see exactly what the engine reports, with nothing softened.
Coaching preview on pre-calibration bands · audio is scored, then deleted immediately and never stored · no emotion inference, ever · 3 runs/hour per person
From waveform to defensible number.
Transcribe
Whisper large-v3 with word-level timestamps and per-word recognition probabilities. Prompted to keep disfluencies it would normally clean up.
Speech geometry
Silero VAD maps exactly when speech happens, including voiced time, silences, and the gaps used for pause classification.
Prosody
Praat pitch tracking, normalized to the speaker's own median in semitones. Energy trajectory catches trailing-off delivery.
Scope & band
Everything is clipped to candidate answer windows, so agent speech never pollutes the numbers. Each feature scores on its trapezoid band.
→Compose
Weighted delivery score with scoring-weight coverage reported. Available evidence uses timestamps in versioned JSON.
One score your tech leads won't argue with.
Delivery informs the score; it never overrides substance. Content, meaning what was actually said and scored for directness, structure, and coherence, carries the majority.
Coverage is the fraction of configured delivery weight that received a score, not the percentage of audio analysed. Withheld measurements remain visible as Not assessed with their returned caveats.
Delivery physics, not mind reading.
Every meter on this page counts something observable in the audio. Nothing infers what a candidate feels. That's a different product category, one we deliberately refuse to build into hiring.
Measured
Countable · observable · replayable- Pace & articulation: articulation rate using content words, with the backend-provided unit
- Response latency & time to first point: read straight off the clock
- Fillers, mid-clause halts, self-repairs: word-level timestamped counts
- Pitch variety: semitones against the speaker's own baseline
- Intelligibility & coverage: model-gated intelligibility and the share of scoring weight assessed
Each one carries a playable clip from the answer itself. Disagree with a number? Press play.
Never
Inferred · unverifiable · excluded by design- Emotion inference: no "nervous," "enthusiastic," or "confident" labels from any modality
- Personality claims: the audio can't tell you who someone is
- Absolute pitch scoring: a deep voice is not a competency
- Accent judgment: intelligibility is measured; accent is not a metric
- Guessed scores: thin or noisy audio results in withheld scores, with coverage reported
This is a product standard, not a regional setting. It holds wherever the interview runs, and it's where regulation is heading anyway.
The numbers only matter if the interview was real.
A parallel pipeline watches for the mechanics of cheating and only the mechanics. It fuses six independent channels on one timeline and produces evidence for human reviewers, never verdicts.
Gaze behavior, not attention
Detects one thing only: the mechanical signature of reading from a second surface, specifically rhythmic left-to-right saccades with line-return sweeps. Looking away while thinking is explicitly normal and never penalized.
Lip-sync verification
Audio-to-lip alignment is calibrated to each session's own baseline offset, so WebRTC's legitimate 100–300ms lag never triggers a flag. Only sustained deviation does; single bad frames never count.
Environment scan
Known assist processes, second displays, virtual cameras, and remote-desktop sessions are checked at the start and logged on the same timeline as everything else.
Typing dynamics
Synthetic input, value injection, and large external pastes are near-conclusive on their own. Softer typing drift is capped per channel so correlated signals can't double-count.
Cross-channel timing
The tell is sequence: focus loss followed by a paste; joint silence followed by a monotone dump; gaze away followed by reading prosody. Acquisition must precede production inside a causal lag window; coincidence doesn't score.
Exculpatory evidence counts
Think-aloud speech overlapping typing, visible revision, and latency that tracks question difficulty all carry negative weight. The system argues both sides before a human sees anything.
Near-conclusive alone: synthetic input events, value injection, a known cheat process, a huge external paste.
Windowed cross-channel pairs: an acquisition event followed by a production event within the causal window.
Soft session-level signals, capped per channel, plus the latency-difficulty gradient and pattern repetition.
Fair questions, direct answers.
Does this penalize non-native English speakers or accents?
Accent is not a metric anywhere in the pipeline. Pitch is scored against each speaker's own baseline, pace against the role's band, and what's measured about clarity is intelligibility: whether the words can be understood, not how they're pronounced. Disfluency counts are language-behavior counts, not accent proxies.
Is this sentiment or emotion analysis?
No, deliberately. Emotion inference in hiring is scientifically weak, and regulators are moving against it. This engine measures delivery physics: things you could count yourself with a stopwatch, a transcript, and a tuner. If a claim can't be replayed as evidence, it doesn't ship.
Isn't gaze tracking just "attention scoring" by another name?
No. The gaze pipeline measures saccade geometry (direction, amplitude, rhythm) and detects exactly one pattern: the reading signature of a second surface. It produces no attention, engagement, or focus scores. Looking away while thinking is explicitly normal in the code's own design contract, and thin or low-confidence tracking returns "inconclusive", never a flag.
Can the integrity system fail a candidate automatically?
Never. It produces evidence timelines, not verdicts. Anything at moderate risk or above routes to a human reviewer with timestamped windows the candidate can review and respond to. Exculpatory evidence (think-aloud, visible revision, difficulty-tracking latency) carries negative weight in the same score.
What audio quality does it need?
Use candidate-only, single-speaker audio where possible. Short or thin samples may produce warnings, measured values with withheld scores, or an input error. Coverage describes scoring weight, not recording duration.
How much does delivery actually count for?
35%. Content (what was actually said, scored for directness, structure, and coherence) carries 55%, and interaction flow the remaining 10%. Delivery informs the score; it never overrides substance.
Can we run this on our own audio?
The analyzer ships with our API for teams that want delivery metrics on their own recordings. ; it's currently gated while we onboard the first partners.
Hear it measure, not guess.
The communication engine and integrity suite run inside every Sharedpro AI interview and include their evidence in every assessment report.
Every measured value carries a playable clip