Self-supervised speech resources for Tigre (ISO 639-3: tig), a Semitic language spoken primarily in Eritrea and Sudan with very limited existing speech-technology support. This repository bundles a Tigre-pretrained HuBERT encoder, a discrete unit-discovery model, forced-aligned transcripts with word-level unit sequences, and a word-to-unit pseudo-lexicon -- everything needed to reproduce or extend this work.

Dataset Summary

6777 audio clips from Mozilla Common Voice (Tigre), each quantized into a HuBERT-derived discrete unit sequence.
6451 clips have full word-level forced alignment (timestamps + per-word unit sub-sequences).
326 clips have a transcript (sentence) but no word-level timing, since forced alignment only ran on the validated.tsv split -- these clips come from other Common Voice splits (train/dev/test/ other/invalidated).
0 clips have neither a transcript nor alignment.
14040 distinct words in the resulting pseudo-lexicon (word -> unit sequences observed across all its occurrences).
A HuBERT-base encoder (12 layers, 768-dim), pretrained from scratch on Tigre audio.
A k-means unit-discovery model (k=100, fit on layer 6 HuBERT features).

How This Was Built

HuBERT pretraining (external to this repo): a HuBERT-base encoder was pretrained from scratch on raw Tigre audio using fairseq, then converted to Hugging Face transformers format.
Unsupervised unit discovery: HuBERT layer-6 features were extracted and clustered with k-means (k=100) to produce a discrete "unit" vocabulary -- no transcripts required for this step, so it covers all 6777 clips.
Forced alignment: Common Voice Tigre clips in the validated.tsv split were aligned to their transcripts using ctc-forced-aligner, built on Meta's MMS model, with romanization to handle Tigre's Ge'ez script. Clips outside validated.tsv still have their transcript pulled in from whichever split they belong to, but without word-level timing.
Joining: word-level timestamps from step 3 were used to slice the frame-level unit sequences from step 2, producing a word <-> unit mapping for every aligned word, plus a corpus-wide pseudo-lexicon.
Validation: word occurrences were checked for consistency between their alignment duration and unit count (flagging cases where a word's time span likely drifted). At full-corpus scale, 0 of ~39,800 scoreable word occurrences were flagged as anomalous, giving reasonable confidence in both the clustering and the alignment quality.

Dataset Structure

data/
  metadata.jsonl   <- one row per clip (see schema below)
  clips/*.mp3      <- audio, straight from Common Voice
pseudo_lexicon.json
hubert-model/      <- HuBERT encoder, standard transformers format
unit-discovery-model/
  kmeans.joblib    <- fitted sklearn MiniBatchKMeans
  config.txt       <- which HuBERT layer + k it was fit with
scripts/
  extract_units.py <- standalone audio -> units example

metadata.jsonl schema (one JSON object per line):

Field	Type	Description
`file_name`	str	Relative path to the audio file (`clips/....mp3`)
`sentence`	str or null	Transcript, if known for this clip (see completeness tiers above)
`units`	str	Whole-clip HuBERT unit sequence (space-separated integers, deduplicated)
`words`	list or null	Per-word `{text, start, end, units_raw, units_deduped}`, only if forced-aligned
`has_alignment`	bool	Whether `words` (word-level timing) is populated -- not whether `sentence` is populated

Loading

from datasets import load_dataset

ds = load_dataset("your-username/tigre-hubert-speech-resources")
print(ds["train"][0])

# Only the fully word-aligned subset:
aligned = ds["train"].filter(lambda r: r["has_alignment"])

Loading the HuBERT encoder + unit-discovery model directly:

from transformers import AutoModel
import joblib

model = AutoModel.from_pretrained("your-username/tigre-hubert-speech-resources", subfolder="hubert-model")
km = joblib.load("kmeans.joblib")  # after downloading unit-discovery-model/kmeans.joblib

See scripts/extract_units.py for a complete, dependency-minimal example of turning a new audio file into a unit sequence with just these two components.

Licensing & Attribution

Audio and transcripts: sourced from Mozilla Common Voice, released under CC0 1.0 (public domain).
This repository's derived content (unit sequences, alignments, lexicon, model weights) is released under CC0 1.0 as well, to keep licensing simple and maximally permissive.
Forced alignment was produced using ctc-forced-aligner, built on Meta's MMS model. The alignment data (timestamps) in this repo are released under this repo's CC0 license; if you plan to use the MMS model itself commercially, check its current license terms separately -- this repo does not redistribute MMS model weights, only alignment output.

Known Limitations

Forced alignment used a zero-shot multilingual model (MMS) via romanization, not a Tigre-specific fine-tuned aligner -- spot-checks looked good, but this is not the same as a dedicated Tigre ASR model.
Only 6451 of 6777 clips have word-level timing (the validated.tsv split); the remaining 326 clips have a transcript but no alignment, and 0 have neither. Check has_alignment before assuming word-level data is present.
The k-means unit vocabulary was fit on a subset of clips, not the full corpus -- units are stable in the anomaly checks performed, but haven't been validated against professional phonetic transcription.
Word-level alignment, not phoneme-level -- units within a word are not further segmented to individual sounds.

Citation

If you use this resource, please cite Common Voice and this repository:

@misc{tigre_hubert_resources,
  title = {Tigre HuBERT Speech Resources},
  year = {2026},
  howpublished = {Hugging Face Hub},
  url = {https://huggingface.co/datasets/your-username/tigre-hubert-speech-resources}
}

Common Voice citation: see https://commonvoice.mozilla.org/en/datasets

Downloads last month: 114

Total file size:

656 MB