Interview transcripts that can be checked against the audio
Producing a transcript is the easy half. The harder half is being able to answer a simple question months later: which words were actually said? Once an interview quote reaches a report, a hiring decision or a published article, a plain text file cannot defend itself. The workable pattern is a review-first routine: generate a working transcript, keep the recording it came from, then verify the passages that carry consequences. Skavio Transcribe covers the transcription step. The checking still belongs to a person.
Checkability is a property of your workflow, not a badge on a tool
A machine transcript is a draft rendering of speech into text. It is useful because it can be reviewed and risky because it looks finished. What makes an interview transcript checkable is not a certification but a set of practical links: the original audio is retained, each transcript is tied to a specific recording, and every correction is recorded with who made it and what the audio actually contained. Skavio Transcribe turns audio into a working text through a web interface; it does not turn that text into a certified record, and an unreviewed output should not be treated as one. Treat the transcript as a claim about the source. The audio is the source.
- Keep the original recording untouched and separate from any processed copy.
- Give every interview a stable identifier that appears in the audio filename, the transcript filename and the review notes.
- Separate the working draft from the reviewed version, and label both so nobody confuses them.
- Record corrections as changes with reasons, not as silent overwrites.
Set up the job so the audio stays findable afterwards
Most failed transcript reviews are filing failures rather than transcription failures. Before processing, make sure the recording is a single file per interview with a name a colleague could match to a calendar entry, a project code or a participant list. Upload through the Skavio Transcribe web page; Skavio's public processing tools are web-based. Pricing runs through the shared Skavio credit wallet, and the figure that matters for a given interview is the server quote returned at submission. That quote, not a remembered rate, is authoritative for the job. Processing uses Groq as the primary engine with AssemblyAI as a fallback and no local speech-recognition fallback in the current runner, so the processing path can differ between jobs. That is a routing detail, not a claim of identical output across engines or formats. Skavio's dated internal measurements — an internal canary on 25 September 2026 handling 20.0 seconds of media in 0.637 seconds of wall time, and a separate run processing 235.01 seconds of audio in 0.98 seconds across 46 segments — describe those specific fixtures only. Actual time depends on file length, format, queue and service availability.
- One interview, one audio file; avoid stitched recordings that mix sessions unless you document the joins.
- Write down the recording date, the participants and the job submission time next to the transcript.
- Note the server quote returned for the job so a later cost discussion is about the actual job.
- Keep the original audio in a location the reviewer can reach — not only on the machine that uploaded it.
The verification pass: check in order of consequence, not from top to bottom
Reading a transcript from beginning to end feels diligent but can miss what matters. A better approach is to audit by consequence. Anything that will be quoted verbatim, used as a number, or attributed as a commitment deserves a replay of the audio. Names of people, companies and products come first, because a wrong name is both obvious and embarrassing. Numbers, dates and amounts come second; spoken figures and regional formats deserve particular attention, and a plausible-looking value is not a check. Negations come third: dropping a single 'not' can invert a sentence while leaving it perfectly readable. Then work through jargon, acronyms, punctuation and anything that looks suspiciously fluent.
- Replay the audio for every passage you intend to quote in public.
- Spot-check proper nouns against an agenda, a signature block or a participant list.
- Confirm numbers, currency labels, dates and percentages by ear, not by plausibility.
- Check for missing or invented negations, and for punctuation that changes how a clause reads.
- Flag overlapping speech and background noise as areas where a human decision is required.
Turn the draft into a record: versioning, sharing and stated limits
A reviewed transcript is a different artifact from the one the service returned, and it should be stored that way. Keep the draft unmodified, keep the reviewed version beside it, and keep the correction log with both. When you share a transcript internally, say which version it is and which passages were checked against the audio. When you share an excerpt externally, include an audio segment reference where the recipient can use one. Consent, privacy and retention questions around recorded interviews need review by whoever owns them; an automated processing tool is not a compliance determination. This workflow should not assume that Skavio Transcribe provides speaker attribution, timing metadata or subtitle export; those outputs are not described in the verified product information. If a review depends on knowing who said what, plan the check around the recording itself and the notes the interviewer took.
- Store the original transcript, the corrected transcript and the correction log together.
- Mark excerpts clearly as quoted or paraphrased before they leave your team.
- Route consent, retention and data-protection questions to whoever owns them rather than settling them inside the transcription step.
- Never distribute a transcript as a verbatim record until the passages that matter have been replayed.
Where human judgement still decides
Transcription output is a draft, and the places that need the closest checking are often the ones interviews turn on: an unfamiliar surname, a specialist term, a number read aloud at speed, two people talking over each other. Skavio Transcribe is built to produce a working transcript for review; it cannot know that a site manager said 'fifteen' and not 'fifty', and it cannot decide whether a half-finished sentence was a commitment or a hesitation. The practical limit of the workflow is this: the transcript narrows the search, and the reviewer settles the question. That division of labour is not a weakness. It is why a checked transcript is worth more than an unchecked one, and why a checkable interview record is a filing habit, an editorial standard and a person replaying a segment, in that order.
- Assume unusual names, technical vocabulary and accented speech will need attention.
- Treat interruptions and crosstalk as unresolved until you hear them.
- Do not accept a fluent sentence as evidence; fluency is not the same thing as a checked quotation.