BeitTigreAI/tigre-hubert-candidate-phones

Draft Candidate Phone Inventory for Tigre — Reviewer Guide

What this is

An automatically-derived candidate sound-unit inventory for Tigre, built from a self-supervised HuBERT model trained on Tigre audio (Common Voice), with sounds grouped by unsupervised clustering (k-means) rather than linguistic analysis.

Two granularities are provided:

• candidate_phones_fine.csv — 100 fine-grained clusters. These likely include allophones (positional/contextual variants of the same phoneme) rather than being one-cluster-per-phoneme.
• candidate_phones_coarse.csv — 38 merged groups, produced by hierarchical clustering of the fine clusters' centroids, targeting the ~35-40 phonemic categories documented in existing linguistic descriptions of Tigre (cf. Raz 1983) and the closely related Tigrinya. This merge is purely acoustic-distance-based and has no linguistic input — treat the grouping as a starting hypothesis, not a conclusion. If your review suggests groups should be split, merged differently, or that the target count of ~38 is wrong for what you're hearing, that's expected and useful information, not something to be reconciled with this notebook's default.

What this is NOT

• Not a validated phoneme inventory. No linguist or native speaker confirmed the categories before this package was built.
• Not aligned to any existing IPA transcription of Tigre.
• Not guaranteed to be clean: a single cluster/group may blend two distinct real sounds, or a real sound may be split across multiple clusters/groups.

How to review

1. Open candidate_phones_coarse.csv in a spreadsheet program — this is the recommended starting granularity, since it's closer to phoneme-scale.
2. For each row, go to the audio_folder column's referenced folder and listen to the exemplar .wav files inside (a handful of short clips per group, each just a real occurrence of that group's sound in context).
3. Fill in IPA_label with your best assessment, confidence_1to5 for how certain you are, and notes for anything ambiguous, split, or merged incorrectly.
4. candidate_phones_fine.csv is available if you want to drop to the finer 100-cluster granularity for any group that looks like it's blending multiple sounds, or note allophone patterns.
5. example_words gives real Tigre words (from forced-aligned Common Voice transcripts) containing that unit/group — useful context, but note this is word-level, not character-level: it tells you the sound occurs somewhere in that word, not which specific letter it corresponds to.

Audio file naming

audio_exemplars/group_XX/clusterYYY_exZ.wav — group XX is the coarse candidate group, clusterYYY is which of the 100 fine clusters this exemplar came from, exZ is just an exemplar number (1 through 6). Each clip is a ~600ms window of real audio centered on one occurrence of that cluster.

Questions / feedback

This package was generated as part of an open Tigre speech-technology project (HuBERT model + Common Voice-based unit discovery). Feedback, corrections, and disagreement with any part of this methodology are genuinely welcome — that's the point of sending this for expert review rather than publishing it as-is.