Browser-based workstation for linguistic fieldwork and comparative analysis. Transcribe recordings, keep timestamps trustworthy, compare forms across speakers, track borrowing evidence, and export clean datasets for downstream analysis.

Status: Active research software. Thesis-critical features land frequently and file contracts still evolve.

Fieldwork data often ends up split across audio editors, spreadsheets, scripts, and notes. That makes it easy to lose track of where a form came from, which timestamp was reviewed, and which comparison decisions are final. PARSE keeps the audio, annotations, comparison table, notes, borrowing evidence, and exports in one workspace.

Fieldwork linguists, comparative and historical linguists, and language-documentation teams working with long recordings, wordlists, and multiple speakers or varieties.

• What PARSE Does
• How PARSE compares to tools you already use
• Research Workflow
• Quick Start
• Annotate
• Compare
• Use with Claude, Codex, and other AI agents
• Documentation
• Research & Citation
• License

• Annotate recordings — listen to long recordings, review each word or phrase, correct IPA and orthography, and keep the timestamps tied to the audio.
• Compare speakers — see the same word or concept across speakers, review variants, and make comparison decisions without rewriting the original data.
• Find words quickly — search across recordings when a target word is buried in hours of audio.
• Track notes and tags — mark forms as confirmed, uncertain, borrowed, dialect-specific, or anything else useful for analysis.
• Check borrowing evidence — bring in contact-language comparison forms with source information and citations.
• Export clean datasets — produce LingPy TSV and NEXUS files for downstream comparative and phylogenetic work.
• Automate repetitive work — use the local API and MCP server when AI agents or scripts need to import speakers, run jobs, or prepare exports.

PARSE does not replace ELAN, Praat, FLEx, or LingPy. It connects them and adds the connective tissue that fieldwork-to-phylogenetics workflows usually require manual stitching to achieve. The fit for someone already using those tools:

What this means in practice

• From ELAN or Praat. You will recognise the waveform-and-tier loop. PARSE has fewer fixed tier types than ELAN but adds a dedicated Compare view and one-click export to LingPy and NEXUS. Spectrograms are there for review, but PARSE is not a phonetics lab — no formant or pitch extraction.
• From FLEx. You will find lighter morphology and lexicon tooling, but stronger support for long fieldwork audio, multi-speaker comparison, and phylogenetic export.
• From LingPy. You will no longer rebuild a Wordlist TSV by hand. PARSE writes one directly from fieldwork audio, with LexStat clustering already run inside the Compare view.

PARSE's role is to be the connecting layer between annotation (ELAN / Praat) and phylogenetics (LingPy / BEAST), with a modern web UI and an agent-driven automation layer that the classic desktop tools do not currently provide.

1. Import or open a speaker recording.
2. Review the audio, text, IPA, and timestamps in Annotate.
3. Compare the same word or concept across speakers.
4. Check borrowing evidence when contact influence matters.
5. Export LingPy TSV or NEXUS for downstream analysis.

For the long-form walkthrough, see the User Guide.

git clone https://github.com/ArdeleanLucas/PARSE.git
cd PARSE
npm install

python3 -m venv .venv
source .venv/bin/activate
pip install -r python/requirements.txt

./scripts/parse-run.sh

Runs on macOS, Linux, and WSL. A GPU is optional; CPU-only works. Open Annotate at http://localhost:5173/ and Compare at http://localhost:5173/compare.

For environment notes, GPU configuration, workspace layout, and troubleshooting, see Getting Started. Adobe Audition marker CSV onboarding is documented in Audition CSV speaker import. For real fieldwork, point PARSE at a workspace root outside the git checkout.

• Review long recordings with a zoomable waveform
• Keep IPA, spelling, meaning, and speaker information together
• Run transcription and alignment help when you need it, then review the result yourself
• Check and correct word boundaries when timing drifts
• Undo mistakes, retime clips, add notes, tag forms, and search for hard-to-find words

Full details in the User Guide.

• Side-by-side table for comparing the same word or concept across speakers
• Review multiple variants for a speaker and choose the form to use for comparison
• Merge, split, or duplicate concept rows for analysis while preserving the underlying source data
• Survey/source labels and color cues for complex elicitation sets
• Tools for grouping related forms and revising those groups as evidence changes
• Borrowing review with contact-language evidence and citation cards
• Shared tags, speaker flags, and export-ready comparison data
• Export to LingPy-compatible TSV and NEXUS for downstream phylogenetic analysis

Full details in the User Guide.

PARSE can also be driven by scripts and AI agents. That means repetitive tasks — importing speakers, running support jobs, checking annotations, and preparing exports — do not have to happen only through the browser. The in-app assistant uses the same bounded PARSE tool layer.

For setup and the full automation reference, see the MCP Guide and Getting Started with External Agents.

Getting Started

• Quick Start
• Installation
• First Full Pipeline Run

Core Concepts

• Architecture Overview — Includes diagrams of the worker system
• How Long-File Processing Works
• Chunking & Subprocess Isolation
• Device Selection (CPU/GPU)

User Guides

• Processing Long Recordings
• Best Practices for Fieldwork
• Troubleshooting Common Issues

Reference

• Environment Variables
• MCP Tools & Job Results
• Job Result Schema
• Configuration Options

For Developers

• Development Guide
• Release Notes
• Architecture Decisions
• Contributing

PARSE was developed for a Southern Kurdish dialect phylogenetics thesis at the University of Bamberg, but the workstation is general-purpose and applies to any fieldwork-driven comparative linguistics workflow involving long recordings, concept-based wordlists, cognate review, borrowing adjudication, and downstream analysis in LingPy, LexStat, or BEAST 2. If you use PARSE in academic work, cite it as research software via CITATION.cff or GitHub's Cite this repository UI.

Ardelean, L. M. (2026). PARSE: Phonetic Analysis & Review Source Explorer [Computer software]. University of Bamberg. https://github.com/ArdeleanLucas/PARSE

See Research Context for full citation guidance and research framing.

MIT License