Speech for trial quality
In a psychiatric or dementia trial, the primary outcome is often produced in a conversation: a rated interview, or a treatment session with a facilitator in the room. When that conversation is recorded, the recording is direct evidence of what the rating and the session were made of.
We read it in two ways. A second reading of the interview, on the participant and on the scale the study already scores. And a measured account of the session: the facilitator's speech, compared between randomised arms. The same pass also returns how the conversation was conducted, read for each interviewer against their own baseline.
What a recorded conversation can tell a sponsor
Both readings use the audio a recorded interview already produces. The scale is administered and scored exactly as before.
A second reading of the rated interview
A consistently derived reading of the participant from the same recorded interview, reported against the endpoint. It says whether the rating moved with the material it was made from.
A measured account of the session
Where a drug is given with psychological support, the facilitator's speech is separated from the participant's and compared between randomised arms, under a test written into the analysis plan.
One capture chain, documented
Device, operating-system build, app and analysis version travel with every file, so a site or a handset is not mistaken for a treatment.
Three layers, one pass over the same recording
Every layer below includes the interviewer. All three come back from one pass over the same audio, in one report, per rater and per interview.
Open or leading
“Since we last spoke”
Rater share of speaking time
Who named a symptom first
Manner after a symptom report
Effort on a slow answer
Interaction: who led the conversation
Read against that rater's own interviews at the same visit and in the same arm, so the patient's stage of treatment is held constant.
Tone: how the rater came across
Each rater compared with their own other interviews.
The flag rule is fixed in writing and set on held-out interviews from the same rater and visit before the read.
Did the rating move, or did the participant?
When entry threshold and endpoint sit on the same instrument, regression to the mean reads as placebo response.
What happened in the session, measured
Blinding questionnaires record guesses. The recording holds what happened in the session.
Prosody and voice quality
Pitch contour, energy and voice quality. A warm and a flat delivery of the same sentence produce the same transcript and different acoustics.
Latency and pauses
Response latency, pause length, overlap and back-channel, measured in seconds from the audio.
Who is speaking
Facilitator and participant separated by voice across sites, without manual annotation, so each side of the session is read on its own.
One package, from capture to read-out
Recording a rated interview or a treatment session is common practice. We package what turns that recording into evidence: the same four parts for both readings, scoped to the protocol.
Capture specification
Recording settings, task timing and consent wording, written into the protocol, with device and app versions logged on every file.
Speakers and quality
Interviewer, facilitator and participant separated by voice, and every recording checked for level, noise and completeness before analysis.
Pre-specified analysis
Features and the comparison drafted for the statistical analysis plan before the data exist: against the scale for interviews, between arms for sessions.
Read-out
Results per arm and per site, each traceable to the recordings and to the analysis version that produced them.
Before the protocol locks, or on audio you already hold
A capture specification and a pre-specified comparison, written into a protocol still in design.
On recordings a completed or ongoing study already holds.
Where would a recording sit in your next protocol?
Tell us what state the recordings are in and whether the protocol is still open.
Get in touch →