Transcription Pro

---
name: transcription-pro
description: Master transcription — speech-to-text workflows, accuracy optimization, speaker diarization, and transcript management.
category: enterprise-communication
---

## Overview

Transcription converts speech to text for meetings, interviews, podcasts, legal proceedings, and accessibility. This skill covers professional transcription practice: choosing approaches (AI vs. human), maximizing accuracy, speaker identification, editing workflows, and managing transcript libraries.


Professional transcription converts audio and video to accurate text: for accessibility, searchability, records, and content repurposing.
Beyond raw conversion, mastery covers accuracy optimization, speaker identification, timestamping, and the workflows that make transcripts genuinely useful.
## When to use

- Transcribing meetings, interviews, or calls
- Choosing between AI and human transcription
- Improving transcript accuracy
- Setting up speaker diarization
- Creating accessible content (captions)
- Building searchable audio archives

- Making video content accessible and searchable
- Creating meeting records and archives
- Repurposing webinars and interviews into written content
- Meeting legal or compliance documentation needs
- Transcribing legal proceedings
## Core concepts

**AI vs. human.** AI transcription: fast, cheap, 90–95% accurate on clear audio — best for searchable archives, rough drafts, and high volume. Human: 99%+ accurate, handles accents/crosstalk/jargon — best for legal, publishing, and critical records. Hybrid (AI draft + human edit) balances cost and quality.

**Accuracy drivers.** Audio quality dominates: close mics, quiet rooms, minimal crosstalk. Then: domain vocabulary (custom vocabularies for jargon/names), speaker separation (fewer overlapping voices), and audio format (uncompressed or high-bitrate). Garbage audio in, garbage transcript out — no model fixes a bad recording.

**Speaker diarization.** "Who spoke when" — essential for multi-person transcripts. Accuracy depends on distinct voices and minimal overlap. Label speakers with real names post-transcription for readability; verify attributions on critical quotes.

**Timestamps.** Word- or sentence-level timestamps enable: click-to-play audio sync, clip extraction, and precise quoting. Essential for legal, research, and content repurposing.

**Editing workflows.** Light edit (fix names, key terms — for internal use), full edit (readability: remove fillers, fix grammar lightly — for publishing), verbatim (every um and false start — for legal/research). Match edit level to purpose; over-editing wastes money.

**Use cases.** Accessibility (captions are often legally required), searchability (find moments in hours of audio), records (decisions, agreements), content repurposing (podcast → blog), and analysis (interview coding, call analytics).


**Accuracy factors.** Audio quality (close mics beat room mics), speaker separation, domain vocabulary (custom dictionaries for jargon), accents and languages, and background noise.
Garbage audio produces garbage transcripts — invest at the source.
Human review for critical content; AI-only for drafts and search.
**Speaker diarization.** Identifying who spoke when — essential for multi-person content.
Enroll voice profiles where possible; label unknown speakers consistently.
Diarization errors compound in long recordings — spot-check.
**Timestamps and formatting.** Word-level timestamps enable search and captioning; paragraph breaks by speaker; filler word handling policies.
Format for the use case: verbatim for legal, cleaned for readability.
Define standards upfront; reformatting later wastes effort.
## Practical workflow

1. **Define requirements.** Purpose (archive? publish? legal?), accuracy needed, turnaround time, speaker count, and budget. These determine AI vs. human vs. hybrid.
2. **Capture good audio.** Separate mics when possible, record locally as backup (not just cloud), test levels beforehand, minimize background noise. This step matters more than model choice.
3. **Transcribe.** Run AI transcription with custom vocabulary (names, jargon, product terms). For critical content, send to human transcription or human-edit the AI draft.
4. **Edit to purpose.** Apply the right edit level. Verify: speaker labels, proper nouns, numbers/dates, and key quotes. Spot-check timestamps on important sections.
5. **Deliver and store.** Format for use (clean transcript, captions file, searchable archive), store audio + transcript together, tag with metadata (date, participants, topic), and set retention per policy.
6. **Leverage.** Search across transcripts, extract clips/quotes, feed into summaries, and analyze patterns (for research or QA).

**Pre-recording checklist:** mics tested → quiet environment → backup recording running → participant names collected (for labels) → custom vocabulary prepared → consent obtained.


**Production workflow:** capture quality audio → transcribe (AI first pass) → human review (critical content) → format per standard → QA spot-check → publish with metadata.
Match effort to purpose — internal search needs less polish than published captions.
**Captioning for video:** accuracy 99%+ for broadcast → speaker labels → sound descriptions [applause] → reading speed (160–180 wpm max) → positioning.
Captions are accessibility infrastructure — treat them as such, not as afterthoughts.
## Common pitfalls

- **Bad audio.** Expecting AI to transcribe a echoey conference room with 8 people on one mic. Fix capture first.
- **Wrong accuracy tier.** Using raw AI output for published/legal content. Match method to stakes.
- **Ignoring speaker labels.** "Speaker 1, Speaker 2" in a 10-person meeting. Label and verify.
- **No custom vocabulary.** Product names and jargon consistently wrong. Provide vocabularies upfront.
- **Over-editing.** Paying for full cleanup of internal archive transcripts. Match edit level to purpose.
- **No timestamps.** Transcripts without time alignment can't link back to audio. Include timestamps by default.
- **Privacy neglect.** Transcripts contain everything said. Access controls, retention policies, and redaction for sensitive content.
- **No audio standards.** Accepting terrible source audio. Publish recording guidelines; reject unusable submissions early.
- **Ignoring turnaround needs.** Real-time captioning vs. post-production have different tools and costs. Match method to deadline honestly.
- **Privacy violations.** Transcribing sensitive conversations without consent. Clear policies on what gets transcribed, stored, and shared.
- **No quality sampling.** Assuming 100% accuracy. Sample and score regularly — accuracy drifts with audio conditions.
- **Ignoring speaker overlap.** Crosstalk producing garbled output. Flag overlapping speech; do not guess attributions.