Workflow leaderboard / Research / Transcription
Turn interview transcripts into coded themes
访谈转写稿主题编码
Produce a traceable first-pass qualitative codebook and coded excerpts from interview transcripts without losing links to speaker turns.
Variants
Free & local / 免费本地
~25 min setup
Any 16 GB laptop is sufficient for transcript coding. Use the separate transcription flagship pack first when starting from audio.
| Component | Role | Price | Link |
|---|---|---|---|
| Taguette 1.5.2 | local qualitative coding and excerpt management | Free / 免费open-source local applicationchecked 2026-08-02 | https://pypi.org/project/taguette/1.5.2/ |
| Ollama 0.32.5 with qwen3:8b | draft codebook and candidate-code suggestions | Free / 免费local runtime; human review remains requiredchecked 2026-08-02 | https://github.com/ollama/ollama/releases/tag/v0.32.5 |
Inductive codebook draft
Act as a qualitative research assistant. Read the supplied speaker-labelled transcript and propose a first-pass inductive codebook. For each code return: short label, operational definition, inclusion rule, exclusion rule, one verbatim example with its existing speaker label and timestamp, and likely overlaps with other codes. Do not infer demographics or motives. Keep contradictory cases. Finish with a list titled 'researcher decisions required'.
Start a private local Taguette workspace
uv venv --python 3.12 .venv
. .venv/bin/activate
uv pip install 'taguette==1.5.2'
taguette
# Taguette opens a browser automatically. If it does not, open http://127.0.0.1:7465, create a project, then import transcript.txt.
Known pitfalls
- LLM-generated codes are suggestions, not analytic findings.
- Do not strip speaker labels or timestamps before import.
- Back up the Taguette project database before changing the code hierarchy.
- A single model pass can flatten minority or negative cases; sample uncoded excerpts manually.
Cloud premium / 云端高配
~30 min setup
Browser and API access only. Use only for transcripts approved for cloud processing.
| Component | Role | Price | Link |
|---|---|---|---|
| AssemblyAI Universal-3.5 Pro Universal-3.5 Pro async API | transcription and speaker-labelled utterances | Usage-based / 按用量计费$0.21 per audio hour plus $0.02 per hour for speaker diarizationchecked 2026-08-02 | https://www.assemblyai.com/pricing |
| Taguette 1.5.2 self-hosted | human-reviewed coding workspace | Free / 免费software is free; hosting cost is separatechecked 2026-08-02 | https://pypi.org/project/taguette/1.5.2/ |
Code application audit
Audit the proposed coding table against the original speaker-labelled transcript. For every coded excerpt verify that the quote is exact, the speaker and timestamp exist, and the operational definition supports the assignment. Return invalid assignments, borderline assignments, uncoded negative cases, and duplicate codes. Do not generate replacement quotations.
Known pitfalls
- Cloud transcription cost varies with audio duration.
- Taguette does not replace consent, anonymization or a data-management plan.
- Avoid treating utterance-level summaries as verbatim evidence.
- Keep a frozen codebook version for each analysis round.