Skip to main content
Flagship case

3,172 interviews, 40+ countries, one pipeline end to end.

Fortune Global 500 · telecoms equipment manufacturer|Global customer satisfaction study

Over one year the client had to hear what customers in 40-odd countries actually said — not scores, but interview recordings, one after another, in 18 languages.

What reached us was audio; what went back was structured findings you can search by theme and drop into a report. Along that pipeline, every major stage is followed by a check, and nothing ships until a person signs off.

By agreement, this page names no client, no product and shows no transcript content. What follows is only counts, structure and the shape of the process; every name in the diagrams is a placeholder.

01

How big this project was

this case · not site-wide

3,172

in-depth interviews · raw material

980k

structured summary rows · final output

438

product terms · count only

599

sub-criteria · the basis for filing

Five dimensions on one ruler

one square = one item · same squares per row, so lengths compare directly

The difficulty here was not duration but dimensions: all 3,172 interviews had to land correctly in all five systems below. Miss in one place and the finding is unusable.

Theme tags1334 top-level groups · 599 sub-criteria · group names are placeholders (A–D)
Customer types44spread across 5 industries · names are placeholders (Type A…)
Countries & regions40+China + 40-odd · across 6 macro-regions · by agreement, unnamed and unmapped
Languages18Chinese + 17 others · one tag system has to hold in all 18
Industries5the 5 industries the 44 customer types belong to · names withheld

Chinese · ES · FR · PT · TH · ID · TR · JA · DE · AR · EN · RU · IT · VI · CS · KO · EL · PL

02

How the pipeline runs

Eight major stages run left to right in one line; each ✓ between them is a check. Stage 07 is human sign-off, item by item — no nod, no output. Click any stage to see its sub-steps.

This pipeline was built for this engagement. What you can run yourself in the product is transcription and review, plus two optional steps — narrative rewrite and redaction; the other stages were part of what we delivered on this project.

scrolls sideways on narrow screens · = check (6 of them)

what this stage does

01Audio → transcript

The same audio runs four times and the four results are held against each other; wherever they disagree gets flagged on the spot rather than guessed at.

audio clean-up and segmentation4 engines transcribing in parallelline-by-line cross-check
03

Why every stage is followed by a check

When a machine rewrites long passages, three problems show up reliably. None of them announces itself — the result usually reads perfectly well, it just no longer matches what was said. So each one has its own checking and correcting stage, compared against the source, corrected on sight.

Problem 1

Invention

The rewrite grows content that was never said — an extra reason, an extra judgement, and it reads perfectly naturally.

what we do about it

Every sentence goes back to the source for warrant; whatever has none is cut.

Problem 2

Loss

Shortening a long sentence drops a condition, a qualifier or a negation — "broadly satisfied" becomes "satisfied", and the finding changes with it.

what we do about it

Checked point by point against the source, and whatever is missing goes back in as it was.

Problem 3

Misattribution

"He", "this one", "over there" attach to the wrong subject — and the sentence now holds someone else responsible.

what we do about it

Re-bound from the surrounding context; anything still uncertain is flagged for human review.

This stage is the heaviest, and the most valuable, part of the whole pipeline.

3,172 transcripts, every one of them all the way through. Claiming accuracy ought to cost us something: if a transcription turns out to have a serious quality defect of our making, that transcription is refunded in full — how to raise it and how we review it is written into §4 of the Terms.

04

What finally goes back

Four deliverables, produced in one run from the same set of transcripts, all agreeing with each other.

01

Interview write-ups

One readable narrative per interview.

.docx

02

Structured summary table

Filed row by row against the theme tags — 980k rows.

.xlsx / .csv

03

Problems and suggestions

What customers themselves raised, collected separately.

web / document

04

Customer-type roll-up

44 customer types side by side, across 5 industries.

.xlsx

Is your project this size? Start by seeing what it costs.

Priced per hour, prorated to the second.