Project Babel: Deciphering a Conlang's Phonology in 48 Hours

This project turns the typically months-long process of phonological analysis into a single-weekend sprint. The target is a moderately complex conlang (e.g., a personal conlang with 20-40 phonemes and simple syllable structure). The researcher arrives Friday night with a laptop, a good microphone, and a willing native speaker. Saturday is dedicated to recording a targeted word list (500-1000 tokens) covering all expected phoneme contrasts, minimal pairs, and syllable positions. Saturday afternoon and evening are for forced alignment (using Montreal Forced Aligner with a rough orthographic transcription if available) or manual segmentation in Praat. Sunday is the core analysis: extracting formant values, VOT, and spectral moments, then running clustering algorithms (k-means, Gaussian mixture models) to propose phoneme categories. The final hours are spent synthesizing an IPA chart and writing up a one-page phoneme inventory with allophonic rules. The entire output is a document that reads like a mini grammar sketch — but produced in under 48 hours.

Most conlangs lack rigorous phonetic documentation; creators often have only a vague idea of their language's sounds. Existing fieldwork methods assume weeks of immersion and multiple consultants. This project tests whether a high-intensity, tool-assisted weekend can produce a credible phonological analysis from a single speaker, potentially accelerating conlang documentation and enabling quick comparisons between planned and actual pronunciation.

• Record a minimum of 500 tokens of a conlang from a single native speaker using a minimal pairs list
• Produce a segmented and time-aligned corpus (TextGrids) covering all tokens
• Extract acoustic features (F1, F2, F3, VOT, duration, spectral centroid) for each segment
• Cluster the segments into a proposed phoneme inventory using unsupervised learning
• Identify at least two allophonic rules (e.g., vowel reduction, consonant lenition)
• Write a one-page IPA summary with rules and an example word list by Sunday 8 PM

• Not a complete grammar — only phonology and phonetics are addressed
• No morphological or syntactic analysis
• Not attempting to model the speaker's internal grammar; only surface acoustic patterns
• No dialectal or sociolinguistic variation (single speaker only)
• Not building a working speech recognition or synthesis system

• Recording: Shure SM57 microphone, Zoom H5 recorder (or laptop with Scarlett 2i2 interface), quiet room
• Segmentation: Praat for manual TextGrid creation; Montreal Forced Aligner (MFA) for automatic alignment if a dictionary exists
• Acoustic analysis: Python 3 with librosa, pyPraat, pandas, scikit-learn; Jupyter Notebook for interactive exploration
• Visualization: Matplotlib, seaborn for formant plots and clustering dendrograms
• Documentation: Markdown, export to PDF via Pandoc

The method is a compressed fieldwork pipeline:

1. Recording (Sat 9-11 AM): Speaker reads from a randomized list of 200 words (each repeated 3-5 times) covering expected phonemes in initial, medial, final positions. Minimal pairs are interspersed.
2. Alignment (Sat 12-6 PM): Use MFA with a hand-crafted pronunciation dictionary (built Friday night) to auto-generate TextGrids. Manually correct any misalignments (expect ~10% error).
3. Feature Extraction (Sat 7-10 PM): For each segmented phone, extract F1-F3 at vowel midpoint, VOT for stops, spectral centroid and skew for fricatives. Store in a pandas DataFrame.
4. Clustering & Analysis (Sun 9 AM-12 PM): Run k-means on normalized feature vectors for each major class (vowels, stops, fricatives separately). Determine optimal k via silhouette score. Propose phoneme labels based on cluster centers and known acoustic targets.
5. Rule Induction (Sun 1-4 PM): Look for contextual patterns: e.g., are vowels laxer in closed syllables? Do stops have aspiration in initial position? List all candidate allophonic rules with evidence.
6. Write-Up (Sun 4-8 PM): Compile IPA chart, phoneme distribution tables, and rule statements. Include spectrograms of key contrasts.

• Speaker fatigue reduces recording quality after 90 minutes — plan breaks
• Insufficient minimal pairs for certain contrasts — may leave some phonemes ambiguous
• Forced alignment fails if pronunciation differs from dictionary — manual correction time balloon
• Clustering may not match traditional phoneme boundaries (e.g., vowel allophones cluster separately) — reinterpretation needed
• Time overrun: feature extraction or clustering may require debugging
• Single weekend constraint means no second pass — results are provisional

• How well does unsupervised clustering recover phonemes known from the conlang's design? (We'll compare with the creator's intended inventory after the weekend.)
• Can effective allophonic rules be inferred from a single speaker's data, or are multiple speakers needed?
• What is the minimal recording time needed for reliable results? Could we shrink to 4 hours?
• Does the weekend timeframe force shortcuts that invalidate results, or is it enough for a useful sketch?

The perfect storm of a willing conlanger with a mature language, a free weekend, and the right equipment never coincided. I moved to a city without a quiet recording space, and the conlanger friend got busy with a new job. The idea remains a tantalizing what-if — maybe next year's Phonetics Hackathon.

This plan assumes the conlang has a romanization or at least a phonemic transcription system. If not, the weekend also includes inventing one. A backup plan: if MFA fails, use a simplified manual segmentation focusing only on 50 carefully chosen words.