Can speech be re‑analysed from the recordings of a completed clinical trial?

Yes. Speech in recordings from a completed clinical trial can be re‑analysed when consent and a legal basis cover further analysis, the audio was kept with its visit dates and scores, and the analysis plan is locked before treatment arms are joined to the readings.

Guide for trial sponsors, clinical operations, biostatistics and digital endpoint teams, by Cephalgo · Updated November 2026

Where the audio is

Where speech already sits in trial data

A trial can hold recorded speech that was never read as speech. Such recordings were made for another purpose, such as rater surveillance or a task in an electronic outcome assessment, and they stay in the trial archive.

Recorded rating interviews

Interviews for scales such as the Clinical Dementia Rating, PANSS and MADRS are often audio‑recorded so that central reviewers can check how each scale was given and scored (Robin et al., 2023; Worthington et al., 2025; Targum and Catania, 2019). The participant speaks at length at every recorded visit. Rating interview QC ›

Speech tasks in eCOA

Where an electronic outcome assessment includes spoken tasks, such as verbal fluency, story recall or picture description, the audio can be stored beside the task score.

Home recordings

Where a study has remote sessions, it may hold short speech tasks recorded on a participant’s own phone or a study phone between visits.

Site‑visit speech tasks

Where a protocol includes motor‑speech tasks, such as a sustained vowel, syllable repetition or reading a passage, they are recorded at the visit.

What to check

What to check before a re‑analysis

Six questions decide whether a set of trial recordings can be re‑analysed, and how much of it. Most can be answered from the protocol, the consent forms and a file listing, before any audio moves.

  1. Consent and legal basis

    Does the consent cover further scientific analysis, and is there a legal basis for it? Under Article 28(2) of the EU Clinical Trials Regulation, consent to use data outside the protocol for scientific purposes can be sought at enrolment, together with trial consent. The European Commission’s view is that such further use needs its own legal basis under the GDPR; the European Data Protection Board adds that the research route of GDPR Articles 5(1)(b) and 89, which treats further scientific use as compatible with the original purpose, should not be ruled out. In the United States, the HIPAA Privacy Rule and the consent and authorization documents set the limits.

  2. De‑identification

    A voice can identify a person. The HIPAA Safe Harbor method lists voice prints among the identifiers to remove, so recorded speech cannot meet Safe Harbor as it is, and is handled by expert determination or under the participant’s authorization, with names, dates and places removed from transcripts and file names, and a data processing agreement in place. Consensus recommendations on reusing trial data cover consent, de‑identification and managed access (Ohmann et al., 2017).

  3. Audio format and quality

    The file format, sample rate and compression, whether the interviewer and participant share one track, how long each recording is, and how many are missing or clipped. A file listing and a small sample answer this.

  4. Recording metadata

    For each file: participant code, visit, date, site, language, and the device and operating system version where they were logged. These link each reading to its visit and show where phones or software changed during the study.

  5. The rated score per visit

    The scale score given at each recorded visit, such as the CDR Sum of Boxes, MMSE or PANSS total, so each speech reading can be placed beside the score from the same visit.

  6. Blinding

    The re‑analysis team builds and locks the model without arm labels; a statistician outside that team joins arms to the locked readings afterwards.

What comes back

What a re‑analysis of trial recordings returns

A re‑analysis turns audio the sponsor already owns into readings that sit beside the trial’s own data, and into numbers the next protocol can use.

  1. Readings on the trial’s own scale

    A reading per recording on the scale the trial used, such as MMSE, CDR or an ALSFRS‑R speech item, at that scale’s own step, joined to the visit score in one table, with the named speech features behind it. What is a speech biomarker ›

  2. Rater‑conduct review

    Where the recordings are rating interviews, a review of how each interview was conducted: length, leading or skipped probes, and score patterns across visits and sites that can point to baseline inflation or rater drift.

  3. Variance components for the next trial

    How much of the variation in readings comes from participants, visits, sites and raters, and an effect‑size estimate computed by the unblinded statistician, so the next trial is sized on the sponsor’s own data.

  4. Phone and software change check

    Where phones or operating systems changed during the study, a check that readings stay stable across the change, with readings recalibrated from the audio where needed. Device changes ›

Design facts

What shapes a re‑analysis

These are properties of recordings made for another purpose. A re‑analysis plan handles each one by design.

The protocol set the tasks

The recordings hold the tasks the original protocol chose. A rating interview yields conversational speech, so the readings follow the scales that conversation can inform.

Consent and legal basis set the sample

Participants not covered by the consent or the legal basis for further analysis are left out, and the analysis states how many recordings were included at each visit.

Two voices on one track

In an interview the rater speaks too. Speaker separation keeps the participant’s speech for the reading and the rater’s speech for the conduct review.

Language and calibration

Each study language is calibrated on data from that language, using the trial’s own screening and baseline ratings.

Published work

Re‑analysis of trial recordings in published work

Peer‑reviewed studies have reused recordings made during clinical trials, to read speech or to re‑score the interviews, and joined the results to the trial’s data.

  1. Alzheimer’s disease: CDR interviews

    Participant speech from Clinical Dementia Rating interviews in a phase 2 prodromal‑to‑mild Alzheimer’s trial gave a speech composite that changed over 18 months with an effect similar to the CDR Sum of Boxes, and correlated with the trial’s primary and secondary endpoints (Robin et al., 2023).

  2. Schizophrenia: PANSS interviews

    PANSS interview recordings collected during a phase 2 schizophrenia trial were processed for speech characteristics, which were associated with positive and negative symptom severity. Recording quality supported timing and rate measures, which is why audio format and quality are checked first (Worthington et al., 2025).

  3. Depression: MADRS interviews

    Audio‑digital recordings of site‑based MADRS interviews from five trials were re‑scored by blinded remote raters and compared with the site scores, as a surveillance method for ratings quality (Targum and Catania, 2019).

Several companies analyse speech for clinical trials; Cephalgo keeps a list of those whose own sites offer it: Speech biomarker companies ›

Regulatory context

How a re‑analysis fits regulatory expectations

A re‑analysis of a completed trial is post hoc, and its place is exploratory: it explains and supports the trial’s findings and informs the design of the next one.

  1. Exploratory analysis

    The ICH E9 guideline notes that data from confirmatory trials are also subjected to exploratory analyses that explain or support their findings and suggest hypotheses for later research. A speech re‑analysis is reported in that frame, with its analysis plan written and dated before arms are joined.

  2. Digital health technologies

    The FDA’s December 2023 guidance on digital health technologies sets out verification, validation and fitness for purpose, and asks sponsors to record the timing and nature of updates, to check that measurements are unaffected, and, where measurements change, to compare data before and after the update.

  3. Qualification

    In the EU, the EMA’s qualification of novel methodologies offers qualification advice on preliminary data and plans, and a qualification opinion on whether a method is acceptable for a specific context of use. Results from a re‑analysis can be part of the preliminary data.

Sources

Guidance, regulation and studies

The requirements and examples on this page come from these documents.

Questions

Re‑analysing trial speech: common questions

Can speech be re‑analysed from the recordings of a completed clinical trial?

Yes, when consent and a legal basis cover further analysis, the audio was kept with its visit dates and scores, and the analysis plan is locked before treatment arms are joined to the readings. Recorded rating interviews, speech tasks in electronic outcome assessments and home recordings can all be re-analysed.

Which trial recordings hold usable speech?

Recorded rating interviews such as the Clinical Dementia Rating, PANSS and MADRS, speech tasks stored in eCOA, motor-speech tasks recorded at site visits, and short home recordings. Rating interviews carry conversational speech at every recorded visit.

Does consent from the original trial allow a speech re‑analysis?

It depends on the consent wording and the law. Under Article 28(2) of the EU Clinical Trials Regulation, consent to use data outside the protocol can be sought at enrolment. The European Commission’s view is that such use needs its own legal basis under the GDPR; the European Data Protection Board adds that the research route of GDPR Articles 5(1)(b) and 89 should not be ruled out. In the United States, HIPAA and the authorization documents set the limits. Participants not covered by the consent or the legal basis are left out.

How is blinding kept in a re‑analysis of a completed trial?

The re-analysis team builds and locks the model on audio and visit scores without arm labels; a statistician outside that team joins arms to the locked readings afterwards.

What does a re‑analysis of trial speech return?

Readings on the trial’s own scale per recording, joined to the visit score with the speech features behind them; a review of how rating interviews were conducted; variance components and an effect-size estimate to size the next trial; and a check that readings stay stable across phone and software changes during the study.

Can regulators use results from a speech re‑analysis?

A re-analysis of a completed trial is post hoc and exploratory. Under ICH E9, exploratory analyses explain or support a trial’s findings and suggest hypotheses for later research. Its results can inform the next protocol and serve as preliminary data in FDA or EMA discussions, such as EMA qualification advice.

Cephalgo

How Cephalgo does this

Cephalgo, a speech biomarker company in Strasbourg, France, offers a fixed‑fee re‑analysis pilot on recordings sponsors already hold, from completed or ongoing trials, where consent covers further analysis. Its multi‑stream model returns a reading per recording on the trial’s own scale, joined to the visit score, calibrated on screening and baseline ratings and locked blind to arm before post‑baseline recordings are read. Readings stay stable across phone and software changes, and where the recordings are rating interviews, Trial QC flags per rater come with them. Fee and timeline are fixed in writing before any data moves. Re‑analysis pilot ›

Holding recordings from a completed trial?

Share the scale, the visit schedule and the consent scope, and get a written scope for a re‑analysis.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.