Media to Transcript

What the buyer receives

A complete, locally processed transcript package for a 45-minute interview recording — five production-ready artifacts, not a raw text dump:

1. Timestamped SRT subtitles

interview.srt — a fully synced subtitle file where every cue is aligned to the audio within ±0.2s. Cue boundaries respect sentence structure and speaker turns, never cutting mid-phrase; overlapping speech is handled with position tags rather than merged lines. SRT indices are sequential and gap-free, so the file imports cleanly into Premiere, DaVinci Resolve, YouTube, and any standard player.

2. Verbatim Markdown transcript with resolved speaker labels

transcript.md — the complete word-for-word transcript, organized with speaker labels resolved against public sources. Instead of generic SPEAKER_00 / SPEAKER_01 tags, the host and guest are identified by name: the pipeline cross-references voice profiles with the episode title, show notes, and public biographies, then confirms identities from named mentions in the conversation itself ("As you wrote in your book..."). Editorial choices are documented inline — filler words are preserved verbatim by default, false starts are marked with an em-dash convention, and a note at the top explains every convention used.

3. Terminology glossary

glossary.md — every domain term, product name, acronym, and proper noun that appeared, each with the canonical spelling and a one-line definition. Built automatically from the title, show metadata, and the decoded transcript, the glossary is what keeps "Kubernetes" from becoming "coober netties" and makes the transcript searchable by the vocabulary the speakers actually used. Reviewers can scan the glossary to verify spellings without re-listening.

4. Per-segment confidence notes

confidence.md — a segment-by-segment quality report. Each section of the transcript carries a confidence rating (high / medium / low) with plain-language notes: "Low confidence at 12:34–12:51 — guest speaks over the host, crosstalk makes attribution uncertain; consider reviewing the source." Background noise, music beds, accented speech, and rapid exchanges are all flagged where they affect accuracy, so the buyer knows exactly which passages deserve a human listen.

5. JSON with word-level timestamps

transcript.json — the full machine-readable transcript with per-word start and end times, per-segment speaker attribution, and confidence scores. Built for downstream editing: buyers can feed it into Descript, clip-search tools, chapter generators, or their own analysis scripts without re-running inference.

How it was made

The 45-minute file was decoded with Qwen3-ASR by default, with automatic fallback to WhisperKit when a segment's confidence dropped below threshold. SpeakerKit diarization separated the two voices before identity resolution. No audio ever left the buyer's machine — every model runs locally, so confidential interviews stay confidential. The whole pipeline, from ingest to the five-file package, runs unattended once the source file is provided.

What makes this different

Most transcription tools return a single flat text file and call it done — no identity resolution, no glossary, no honesty about where the model struggled. This deliverable is a reviewed artifact: labeled speakers, verified spellings, flagged uncertainties, and an edit-ready JSON layer. The buyer gets the SRT they asked for plus everything a serious editor needs downstream.