Datasetpublic

mmwanje/waxal-features-v1

waxal-features-v1 Precomputed Whisper-large-v3 log-mel input_features + tokenized labels for Google WaxalNLP (Lingala, Shona, Luganda). Use this to skip FLAC download + feature extraction when fine-tuning openai/whisper-large-v3 (or any model that consumes the same Whisper-v3 mel / tokenizer layout). Contents Field Type Notes id string Clip id language string lin / sna / lug split string Source split tag input_features list[list[float16]]… See the full description on the dataset page: https://huggingface.co/datasets/mmwanje/waxal-features-v1.

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes248downloads

Dataset Card

waxal-features-v1

Precomputed Whisper-large-v3 log-mel input_features + tokenized labels for Google WaxalNLP (Lingala, Shona, Luganda).

Use this to skip FLAC download + feature extraction when fine-tuning openai/whisper-large-v3 (or any model that consumes the same Whisper-v3 mel / tokenizer layout).

Field	Type	Notes
`id`	string	Clip id
`language`	string	`lin` / `sna` / `lug`
`split`	string	Source split tag
`input_features`	list[list[float16]]	Whisper-v3 log-mel
`labels`	list[int64]	Token ids without language prefix tokens

—38 198 examples · ~59 parquet shards · ~29 GB download
—Recipe id: whisper-v3proc_nolang_f16_v1

Load

python

from datasets import load_dataset

ds = load_dataset("mmwanje/waxal-features-v1", split="train")
# each row: id, language, split, input_features, labels

With the WAXAL model.py pipeline:

bash

export WAXAL_FEATURES_DATASET=mmwanje/waxal-features-v1
python model.py --stage train --model_id openai/whisper-large-v3 --aug_mode spec

Not for Omni

These features are Whisper-v3 specific. Omnilingual ASR needs `mmwanje/waxal-omni-parquet` instead.