demeleww/Amharic_tokens

OmniVoice Amharic v4 — Pre-extracted Audio Tokens

Pre-extracted HiggsAudioV2 audio tokens for training OmniVoice Amharic TTS.

Downloading these tokens (~648 MB, ~2 min) skips the 30+ minute token extraction step (steps 1-3 of the training pipeline).

Dataset Stats

Stat	Value
Total samples	81,731
Total audio	~331 hours
Avg sample duration	~14.6 seconds
Audio shards	164 (.tar files)
Text shards	164 (.jsonl files)
Samples per shard	500 (last shard: 231)
Total size	~648 MB
Tokenizer	eustlb/higgs-audio-v2-tokenizer
Codebooks	8 (vocab size: 1025, mask_id: 1024)

Source Datasets

Dataset	Description
google/WaxalNLP (amh_asr)	Largest source — Amharic ASR data
gheero-Leyu/leyu-amharic-addis-ababa-dialect	Addis Ababa dialect only (other dialects excluded to avoid accent issues)
surafelabebe/amharic_clear_audio_tts	Clean TTS-oriented recordings
chappM/amharic-bdu-asr	BDU ASR dataset

Filters Applied

• Duration: 1–25 seconds
• Text length: max 300 characters
• Script: Ge'ez (ግዕዝ) script only — Latin transliterations filtered out
• Dialect: Only Addis Ababa from Leyu dataset (Gojjam/Wello/Shewa/Gonder excluded)

File Structure

├── data.lst                    # Manifest: paths, sample counts, durations per shard
├── audios/
│   ├── shard-000000.tar        # WebDataset tar: tokenized audio arrays
│   ├── shard-000001.tar
│   ├── ...
│   └── shard-000163.tar
└── txts/
    ├── shard-000000.jsonl      # Matching text/metadata for each audio shard
    ├── shard-000001.jsonl
    ├── ...
    └── shard-000163.jsonl

data.lst format

Each line: <audio_tar_path> <text_jsonl_path> <num_samples> <total_duration_seconds>

Note: The paths in data.lst reference the original Colab extraction path (/content/omnivoice_amharic_v4/wds_tokens/...). The training script rewrites these to local paths after download.

Usage

These tokens are automatically downloaded by train_waxal_v4.py:

from huggingface_hub import snapshot_download

# Download all tokens (~648MB)
snapshot_download(
    repo_id="demeleww/omnivoice-amharic-v4-tokens",
    repo_type="dataset",
    local_dir="/content/omnivoice_amharic_v4/wds_tokens",
    ignore_patterns=[".gitattributes", "README.md"],
)

Or run the full training pipeline which handles this automatically:

!pip install -q omnivoice webdataset tqdm huggingface-hub
from huggingface_hub import login
login()
!wget -O /content/train_waxal_v4.py "https://huggingface.co/demeleww/omnivoice-amharic-training/resolve/main/train_waxal_v4.py"
%run /content/train_waxal_v4.py

Related

• Training script: demeleww/omnivoice-amharic-training
• Base model: k2-fsa/OmniVoice
• OmniVoice paper: arxiv 2604.00688
• Georgian recipe reference: NMikka/omnivoice-finetuned-ka