transcript-cleaner

Fix messy interview transcripts in seconds — speaker labels, filler words, fragmented turns, and broken timestamps. Rule-based by default. LLM-powered when you need it.

---

The problem

Every UX researcher knows this pain. You finish a one-hour interview, export the transcript from Zoom, Otter, or Rev — and the file is a mess:

• Kostas K, K Kazakos, and interviewer are all the same person
• Half the turns start with "um, like, you know, sort of..."
• Three consecutive "Yeah." turns that should be one
• Timestamps in five different formats across the same file
• Turns with no speaker label at all

Fixing this manually takes 1–2 hours per transcript. This tool does it in seconds.

---

Features

Operation	Approach
Fix speaker label inconsistencies	Rule-based normalisation + fuzzy matching + optional LLM
Fill unlabelled / silent turns	Heuristic continuation + optional LLM context inference
Remove filler words	Regex pipeline (um, uh, like, you know, sort of, basically...)
Merge fragmented short turns	Same-speaker merge with configurable word threshold
Normalise timestamps	Handles HH:MM:SS.mmm, M:SS, float seconds, plain integers

Model-agnostic — works with OpenAI, Anthropic (Claude), or Google Gemini. No LLM required for most operations.

Supports all major transcript formats:

• Plain .txt (manual formatting)
• Zoom / Microsoft Teams auto-transcripts
• Rev / Trint .json
• Otter.ai .txt and .docx

---

Installation

# Minimal (rule-based only, no LLM)
pip install transcript-cleaner[cli]

# With a specific LLM backend
pip install "transcript-cleaner[cli,anthropic]"
pip install "transcript-cleaner[cli,openai]"
pip install "transcript-cleaner[cli,gemini]"

# Everything
pip install "transcript-cleaner[all]"

---

Quick start

CLI

# Rule-based only — no API key needed
transcript-cleaner clean my_interview.txt

# With LLM (Claude) for speaker resolution and unlabelled turns
transcript-cleaner clean my_interview.txt --llm anthropic --model claude-sonnet-4-5

# Batch clean a whole folder
transcript-cleaner batch interviews/*.txt --output-dir cleaned/ --llm openai

# Dry run — see the report without writing output
transcript-cleaner clean my_interview.txt --dry-run

Example output:

📄  Parsing  session_04.txt ...
    Format detected: zoom  (47 turns)
🔧  Running cleaning pipeline ...

✅  Cleaning report:
  Turns:              47 → 39
  Speaker fixes:      12
  Unlabelled filled:  3
  Fillers removed:    28
  Turns merged:       8
  Timestamps fixed:   47
  LLM calls made:     0

💾  Saved → session_04_cleaned.txt

Python API

from transcript_cleaner import parse, run_pipeline, TranscriptFormat

# Parse
transcript = parse("session_04.txt")
print(f"Format: {transcript.source_format.value}, Turns: {len(transcript)}")

# Clean (rule-based)
cleaned, report = run_pipeline(transcript)
print(report.summary())

# Clean with LLM
from transcript_cleaner.llm import get_backend
llm = get_backend("anthropic")  # reads ANTHROPIC_API_KEY from env
cleaned, report = run_pipeline(transcript, llm=llm)

# Export
print(cleaned.to_text())

---

Format support details

Plain .txt

Expects Speaker Name: turn text format, with optional [HH:MM:SS] prefix.

Zoom / Teams

Standard auto-transcript export: timestamp on one line, speaker name on the next, then text block.

Rev / Trint .json

Supports Rev's monologues format and the generic AWS Transcribe / Trint segment format.

Otter.ai

Both .txt export (speaker name → timestamp → text blocks) and .docx export (requires pip install "transcript-cleaner[docx]").

---

Configuration

All flags can be combined:

transcript-cleaner clean interview.txt \
  --llm anthropic \
  --model claude-haiku-4-5 \
  --no-fillers \          # keep filler words
  --no-merge \            # keep short turns separate
  --json-report           # machine-readable report

---

Contributing

Bug reports, format support requests, and PRs welcome. If you use a transcript format not listed above, open an issue with a sample (anonymised) — I'll add a parser.

---

Part of the CPR Ecosystem

transcript-cleaner is a satellite tool in the Computational Product Research (CPR) ecosystem — a discipline building AI-augmented infrastructure for rigorous, human-centred product research.

Related projects:

• CausalTrack — detect the Say-Do Gap in user research sessions
• synthetic-user-council — generate richly profiled synthetic research participants

Learn more about CPR at github.com/kgkazakos.

---

Citation

If you use this tool in research, please cite:

@software{kazakos2026transcriptcleaner,
  author  = {Kazakos, Kostas},
  title   = {transcript-cleaner: Rule-based and LLM-powered transcript cleaning for UX research},
  year    = {2026},
  url     = {https://github.com/kgkazakos/transcript-cleaner}
}