Research Engine — user guide

The Research Engine turns a recording into a written transcript, works out who said what, and removes personal details — names, phone numbers, addresses and so on — replacing each one with a placeholder like [PERSON_1]. You can then read it against the audio, correct it, ask the AI about it, search across all your transcripts, look for common threads across several calls at once, and share the result with colleagues. Everything runs on our own machines; recordings and transcripts never leave the building.

1. The workflow at a glance

  1. Upload a recording into one of your projects and choose what should be redacted.
  2. Wait while the app transcribes, identifies speakers, double-checks the words with a second listener, and redacts personal details. This is automatic.
  3. Review — work through the short list of lines the two listeners heard differently (most are settled for you), and skim the redactions list.
  4. Use it — read the redacted transcript alongside the audio, correct anything that's wrong, ask the AI about it, search across your whole library, or line several calls up side by side to find common threads.
  5. Share the job with a colleague, or download the redacted text and a redacted audio version.

2. Signing in

The sign-in screen with name, password and 2FA code fields
  1. Your name or email address — either works.
  2. Your password. On your very first sign-in this is the temporary password from your invitation email; the app will make you choose a new one before you can do anything else.
  3. The current 6-digit code from the authenticator app on your phone. Your invitation email includes a QR code — scan it with any authenticator app (Google Authenticator, Microsoft Authenticator, …) to set this up.
  4. Press Login.

Once you're in, the header shows “Signed in as …” so you always know which account you're using. Sessions expire after several hours — if the app suddenly asks you to sign in again, that's all it is. Nothing is lost; jobs keep running while you're signed out.

New phone? After signing in, open Settings → Password / 2FA and press Show my 2FA QR code to scan it again.

3. Uploading a recording

Uploads live inside your projects (see section 5). In the library on the left, open the project folder the recording belongs in, find its Transcripts sub-folder and press + Add, then pick an audio or video file — the sub-folder opens by itself so the new job is in view. Before anything is sent, the app asks what you want redacted:

The redaction options dialog shown before an upload starts
  1. Redact personal details — untick this only if you want a plain transcript with nothing removed (for a recording that contains nothing sensitive).
  2. What to redact — the kinds of personal detail the app will look for. They're all on by default; untick any you want to keep in the transcript.
  3. Upload + process starts the job. Cancel abandons the upload; nothing is sent.

Long recordings are fine — files upload in pieces, and an hour-plus meeting is normal. Each upload becomes one job in that project's Transcripts sub-folder, named after the file (you can rename it later with Edit). The Files and URLs sub-folders alongside it take written sources — PDF documents and web pages — described in section 5.

4. While it processes

A finished job expanded in the library, showing its COMPLETE status chip and the Edit, Move, Share and Delete buttons — the same row shows a bar per stage while it is still processing
  1. The job's status. A finished job reads complete; while it is still working the same chip shows processing and a bar appears for each stage — Transcribing audio (speech to text), Identifying speakers, Comparing transcript (a second engine listens again and flags lines it heard differently) and Redacting personal details — turning to failed with a message if something goes wrong.
  2. Edit renames the job or adds notes, Move refiles it into a different project, Share gives another user access, Delete removes the job and all its files.

Everything runs on the server — you can close the browser and come back later. The library updates itself every few seconds while a job is running, and the sub-folder holding a running job opens by itself so the bars are in view. As a rough guide the transcription stage takes about a quarter of the recording's length; redaction adds a few more minutes.

5. Your library

The left-hand panel is your library: your project folders, with everything belonging to each piece of work — its sources and its cross-references — kept together inside one folder. Above the list is a Search all transcripts box that searches inside every transcript you have (see section 11).

Every account starts with a General project, and + New project at the top creates more — one project per piece of research is the idea, so a cross-reference and the transcripts it draws on sit side by side in the same folder. Click a project's name to open it: inside are four sub-folders, each showing a count of what it holds — Files (PDF documents), Transcripts (audio and video recordings), URLs (web pages) and Cross-references. The three source sub-folders each have their own + Add; Cross-references has + New instead (see section 12). Sub-folders start closed, so a project reads like a file system rather than one long list — click a sub-folder to open it. Two things open one for you: a job still processing inside it (so its progress bars are visible), and starting an + Add, which opens the sub-folder the new job will land in. A project row also offers Rename, and Delete once the project is empty. Sources and cross-references alike carry a Move button — the same project picker for both — to refile them into a different project as your filing evolves. Anything colleagues have shared with you — recordings, documents and cross-references alike — appears in its own Shared with me group at the bottom, never mixed into your projects.

Adding documents: PDFs and web pages

Files → + Add takes a PDF. Before the upload starts the app asks one question: “Reference material — contains no personal information”. Leave it unticked (the default) and personal details are redacted from the extracted text, exactly as they would be for a recording — you get an original and a redacted copy. Tick it only for public or reference documents: the text is then kept as written, with a single copy and nothing removed. Scanned/image-only PDFs can't be read — you'll get a message asking for a text PDF instead.

URLs → + Add takes a web page address. Public https:// pages only — the page is fetched once and stored; it is never re-visited, so add it again if you want a fresher copy. The readable article text is extracted and the job appears under URLs, named after the page. Pages behind logins, internal addresses and non-HTML links (a direct PDF link belongs under Files) are refused with an explanation.

A finished document job offers its extracted text (Document text for a PDF — with an Original PDF download alongside it — or Snapshotted text for a web page). Opened in the viewer it reads like a transcript, but each paragraph is labelled with its page and section (e.g. Page 3 — Methods) instead of a time and speaker, and there is no audio player — nothing was recorded. Search, the ✎ text correction, Ask AI (section 10) and the library-wide search all work on documents just as they do on transcripts; sharing works with the same Redacted / Original document sets (a reference document or web page has only its one copy, so either grant reveals it).

A project open in the library showing its four closed sub-folders — Files, Transcripts, URLs and Cross-references — each with a count and its own Add button
  1. Click a project to open it and its four sub-folders appear — Files, Transcripts, URLs and Cross-references — each with a count of what it holds. They start closed, so a project reads like a file system; click a sub-folder to open it.
  2. Each source sub-folder has its own + Add; Cross-references offers + New instead.

You only see your own jobs, plus any that colleagues have shared with you (marked “Shared with you”). What you can see and do inside a shared job depends on what the owner granted you — see section 13.

6. Reading a transcript

The transcript viewer with the audio player, search box and a highlighted word
  1. The audio player. While it plays, the transcript follows along, highlighting the word being spoken and scrolling to keep it in view.
  2. Click any word to jump the audio to that exact moment — the fastest way to check what was really said. The timestamp button at the start of a line does the same for the whole line.
  3. Search the transcript; Prev / Next step through the matches. (While a search is active the word-by-word highlight pauses — clear the box to get it back.)
  4. Each line shows its speaker. Click the speaker name to correct it — you can reassign the whole line, or split it where a new person starts talking. The ✎ button at the end of a line lets you correct the words themselves.

7. The accuracy review

Two different speech engines listen to every recording. Where they disagree, the line is flagged amber in the transcript with a note showing what the second engine heard. An AI pass settles the obvious ones for you (those show green, with the reasoning) — what's left is genuinely ambiguous and worth human ears.

A flagged line in the transcript with the cross-check engine's alternative underneath
  1. A flagged line — the amber note underneath shows what the cross-check engine heard instead. Click the note to settle it in the review dialog. Notes with a green ✓ are already settled (by you or the AI) and can be clicked to reopen.

Press Review in the toolbar to work through everything still open:

The accuracy review dialog comparing what the two engines heard
  1. What each engine heard, with the differing words highlighted.
  2. ▶ Play line plays just this moment of the recording (it starts a second early so you get a run-up).
  3. Pick a side: Whisper is right keeps the main transcription, Cross-check is right takes the second engine's words, Both wrong — edit lets you type the correction yourself.
  4. Prev / Skip move through the queue without deciding.

Some disagreements matter far more than others, and the review marks those for you. When the two engines heard a different number, a not/no (a negation) or a name, the line gets an amber ⚠ number, ⚠ negation or ⚠ name chip — even when the rest of the line matches word for word. A single swapped number, or a missing “not”, can change what a call means, so these are the ones most worth putting your ears on. A PII chip marks a flag whose line also carries a redacted personal detail.

Each flag also shows how sure the main engine was of its own words — for example “whisper 82% sure”. The queue is ordered to bring the riskiest lines to the top: the critical ⚠ disagreements first, then the ones the engine was least confident about, with anything touching redacted personal detail moved up — so the time you spend listening lands where it counts.

The AI does three jobs with the flags before you ever open the review: it dismisses trivial differences (a spelling, a filler word), corrects the clear-cut mistakes, and — for anything still unsettled — brings in a third speech engine (the “arbiter”) to listen to that exact moment again. When two of the three engines agree, the line is settled for you; when the arbiter still can't decide, what it heard is shown as a third row of evidence in the review, so you have three independent readings to weigh up. Every one of these decisions is only a starting point — you read the evidence and can override any of them.

Decisions are saved instantly and the transcript regenerates around them — including re-redacting any line you change. Already-reviewed lines (yours or the AI's) can be reopened by clicking their green note in the transcript, so nothing is ever final. If the same mistake appears several times, the app offers to fix all of them in one go. When a job is shared for editing, colleagues work the same queue together and each decision records who made it.

Teaching the app rare words

Every workplace has words the app won't know at first — a product name, a place, some jargon, an unusual surname. When you correct one of these while reviewing, the app quietly remembers it and nudges your future transcriptions (and the AI) toward that spelling, so you don't have to keep fixing the same word.

You can see and manage this list yourself under Settings → Glossary: add a term by hand, or remove one you don't want. It's your own list — it only shapes your own future recordings, and it isn't shared with anyone.

8. Editing a transcript

Beyond the accuracy review, you can correct anything in the transcript directly:

Every change is recorded against your name and the time you made it, so on a shared job you can always see who edited what. If two people edit the same job at once and the transcript moves under you, a save can come back with “transcript changed” — just reload and redo that one edit (see section 16).

Editing is available to the owner and to anyone the owner has shared the job with as an Editor. Everyone else can read and listen but not change anything.

9. Checking the redactions

With the redacted transcript open, press Redactions in the toolbar to see everything that was removed:

The redactions list, grouped by person and organisation
  1. One entry per person or organisation — every way their name was said (“John Davis”, “John”, “Mr Davis”) shares a single placeholder, listed together with all its occurrences underneath.
  2. Each occurrence shows when it was said — click the row and the audio jumps there so you can verify the redaction against the recording.
  3. Unredact puts the original words back into the redacted transcript for that one occurrence (after a yes/no confirm) — for the times the app was overcautious. The change is recorded, and a Redact button on the row reverses it.

This list shows the original details behind the placeholders, so it's only open to the job's owner and to colleagues the owner has given the Original (un-redacted) set. Anyone shared only the redacted version can't see what sits behind a placeholder.

10. Asking the AI about a transcript

With a transcript open, the AI button at the right of the toolbar lets you ask questions about it in plain English — "summarise this call", "what did they agree?", "list every action point". The AI runs on our own machines, like everything else; the transcript is never sent to an outside service.

The AI button at the right end of the toolbar
At the right end of the toolbar, next to Redactions.
The Ask about this transcript dialog with a question and its answer
  1. Type your question. It's answered from the transcript you have open — ask from the redacted version and the AI only ever sees the redacted text.
  2. Ask AI sends it. The local model is thorough but not fast — the panel shows a running counter, and a long answer can take a few minutes. You can close the panel and carry on; the answer appears when it's ready.
  3. The answer, quoting the transcript where relevant. Your recent questions and answers are kept below so you can look back at them. A saved question can be removed with the Delete button beside it — your own always, or any of them if the transcript is yours.

Ask AI is available to anyone a job is shared with, answering from whichever version (redacted or original) they're allowed to see.

12. Cross-referencing several calls

Cross-referencing lines several finished calls up together and finds the threads that run through all of them — the recurring topics, the common complaints, the points that keep coming up — and writes them up for you, with quotes and links back to where each one was said. It's the tool for a question like “what are the main themes across this month's complaint calls?” The work is done by the AI on our own machines; it's thorough rather than quick, so a set of long calls can take a while and runs in the background.

An analysis can also mix calls with written sources — uploaded PDFs and saved web pages sit in the same picker, marked with a document chip. A document has no recording to check against, so it needs no accuracy review: once it has finished processing it's ready to include. Quotes taken from a document cite its page and section (e.g. Page 3 — Methods) instead of a moment in a call.

A project's Cross-references sub-folder open in the library, with a New button and a Move button on a saved cross-reference
  1. Open a project and find its Cross-references sub-folder, alongside its Files, Transcripts and URLs. Your saved cross-references for that project are listed inside; click one to reopen it.
  2. + New starts a new cross-reference for that project.
  3. Move refiles a cross-reference into a different one of your projects — just like moving a source — from its row in the tree or the button on its detail bar; its member transcripts stay where they are.

Setting one up:

The New cross-reference form: a project selector, a name field, an Original/Redacted choice and a list of eligible transcripts to tick
  1. Choose the project the cross-reference belongs to. Your own calls listed for picking come from that project, under This project; calls shared with you are offered too, under Shared with me, whichever project you choose.
  2. Give it a name you'll recognise later, e.g. “Complaint calls June”.
  3. Choose whether to work from the Original transcripts or the Redacted ones. Use redacted if the result might be shared more widely.
  4. Tick two or more calls. Only fully-reviewed transcripts can be included — the job must be finished and every accuracy flag settled — so the analysis is built on words you've confirmed. A call that isn't ready yet is greyed out, with the reason shown beside it.
  5. Set it going. It runs in the background; you can leave the page and come back. If it's ever interrupted it can be picked up again with Resume.

When it finishes you get a write-up with a few different views to explore, along the top:

A finished cross-reference showing the Report view and the row of view tabs above it
  1. Report — the main write-up: the themes that run across the calls, what each one is about, and how often and where it came up.
  2. Evidence wall — the actual quotes behind a theme, so you can see exactly what was said; click a quote to jump to that moment in that call.
  3. Theme map — a coloured overview of how the themes relate and how strongly each call features them.
  4. Fingerprints — a per-call summary, showing which themes each individual call touched on.
  5. Questions — ask a question across the whole set (see below).

The Report view tells you not just what the themes are but how solid each one is:

Every quote is the call's own words, verbatim, and clicking one still jumps you straight to that moment in that call. To keep or pass on the whole analysis, press Download report (on the open cross-reference): it saves the entire cross-reference — themes, quotes and all — as a single self-contained web page you can open in any browser, print, or “Save as PDF”.

The Questions view lets you ask about the whole set at once — “which calls mention a refund?”, “what did customers most want changed?” — answered from the analysis, with pointers back to the calls involved:

The Questions view of a cross-reference with a question typed in and an answer drawn from across the calls
  1. Type a question across these transcripts and press Ask. As with the single-transcript AI, a full answer can take a minute or two.
  2. The answer, drawing on all the calls in the set at once. Each saved question carries a Delete button beside it — remove your own, or any of them if you own the cross-reference.

You can change which calls are in a cross-reference after it's built. The Transcripts button (on the open cross-reference, beside Delete) opens the list of calls: untick one to drop it, tick another fully-reviewed call to add it — two to forty in all — then Save & re-map. It rebuilds the write-up around the new set, reusing the work already done on the calls you kept and only reading the ones you've added — so it's far quicker than the first run — and it refreshes any saved questions so the whole analysis stays in step. Only the owner can change the set, and not while it's still running.

There's a sharing side to removing a call: it also takes that call away from the people you shared the cross-reference with — unless the call is their own, or you shared it with them separately. Adding a call gives them the new call too, so they can see the evidence behind the fresh themes.

If you later edit one of the included calls, the cross-reference notices it's out of date and offers an Update to fold the change in — it only re-does the calls that actually changed, so it's quicker than starting over. Delete removes the cross-reference (the calls themselves are untouched), and Share lets a colleague see it, the same way you share a job.

13. Sharing with colleagues

The Share button on a job (and on a cross-reference) lets you give chosen colleagues access. Sharing is decided per person: for each colleague you choose which version they can see and whether they can edit.

The Share dialog with a colleague ticked and the Redacted / Original / Editor options beside them
  1. Tick a colleague to give them access; untick to take it away.
  2. Redacted — they can see the redacted transcript and the redacted audio. This is the safe default: personal details stay hidden behind placeholders.
  3. Original (un-redacted) — they can also see the original transcript and the source recording, and the list of what was redacted. Only give this to people who are allowed to see the personal details. The two boxes are independent — you can grant one, the other, or both.
  4. Editor (can review/edit) — they can work the accuracy review and correct the transcript, not just read it. Editing works on the original words, so ticking Editor includes the Original set automatically.
  5. Save sharing applies it. Colleagues then see the job in their own library marked “Shared with you”.

A newly-added colleague starts with the redacted version only — you add the original set or editing deliberately. Whatever you grant, managing the job stays with you: renaming, deleting, sharing it onward and uploading are the owner's alone. Anyone you share with can use Ask AI on the version they can see.

14. A redacted audio version

As well as the redacted text, you can create a redacted audio version of the call — the transcript read aloud by generic voices, with a short bleep in place of every redaction, so there's a shareable recording that gives nothing away. Open the redacted transcript (the redacted_transcript.json view) and find the button in the audio bar:

The Generate redacted MP3 button in the audio bar
In the audio bar above the transcript. It's only offered on the redacted transcript. Once an audio version exists the button reads 🔊 Regenerate redacted MP3.

Pressing it opens a short dialog to choose a voice for each speaker:

The redacted-audio dialog with a Male/Female choice per speaker
  1. One row per speaker — pick a Male or Female voice for each. A distinct generic voice is used per speaker; the real voices are never used, and every redaction becomes a short censor bleep.
  2. Generate audio starts it. This runs in the background (a few minutes for a long call) — you can carry on working or close the dialog.

When it's ready the recording appears as Redacted audio in the job's Audio list on the left, ready to play or download like any other file:

The finished Redacted audio entry in the job's file list
  1. The finished Redacted audio — click it to play. If you later edit the transcript, press Regenerate redacted MP3 to refresh it.

15. Getting the result out

16. If something looks wrong