mmwanje/waxal-whisper-large-v3-spec_aug_v1
Automatic Speech Recognition • 2B • Updated • 44
Datasets:
id stringlengths 5 10 | language stringclasses 1
value | split stringclasses 2
values | input_features listlengths 128 128 | labels listlengths 8 184 |
|---|---|---|---|---|
lin_1 | lin | train | [[-0.56201171875,-0.56201171875,-0.56201171875,-0.56201171875,-0.56201171875,-0.56201171875,-0.56201(...TRUNCATED) | [50258,50364,629,8588,2312,277,8308,20151,5103,5207,4037,22854,9869,297,656,1301,1714,64,11,20151,29(...TRUNCATED) |
lin_100012 | lin | train | [[-0.646484375,-0.646484375,-0.646484375,-0.646484375,-0.646484375,-0.646484375,-0.646484375,-0.6464(...TRUNCATED) | [50258,50364,32,4151,520,8835,17049,2194,13599,275,897,1220,846,13372,7362,84,308,2394,2533,72,308,2(...TRUNCATED) |
lin_100015 | lin | train | [[-0.57177734375,-0.57177734375,-0.57177734375,-0.57177734375,-0.2919921875,-0.389404296875,-0.50146(...TRUNCATED) | [50258,50364,32,4151,20151,298,4037,1667,45610,8588,13704,2478,17818,654,308,2394,43578,4711,262,986(...TRUNCATED) |
lin_100016 | lin | train | [[-0.62353515625,-0.62353515625,-0.62353515625,-0.62353515625,-0.62353515625,-0.62353515625,-0.62353(...TRUNCATED) | [50258,50364,33824,19176,277,8308,8835,17049,3119,514,8021,297,3680,14366,13704,2478,220,273,34241,2(...TRUNCATED) |
lin_100018 | lin | train | [[-0.6396484375,-0.6396484375,-0.6396484375,-0.6396484375,-0.6396484375,-0.6396484375,-0.6396484375,(...TRUNCATED) | [50258,50364,44,897,5103,277,8308,987,304,7849,13659,5159,2478,2453,1123,27147,78,4023,64,262,1035,2(...TRUNCATED) |
lin_10003 | lin | train | [[-0.9091796875,-0.9091796875,-0.9091796875,-0.9091796875,-0.9091796875,-0.9091796875,-0.9091796875,(...TRUNCATED) | [50258,50364,44,34922,2478,43029,556,64,8588,23337,11,308,2394,1667,2265,72,2478,47690,650,13,220,36(...TRUNCATED) |
lin_100031 | lin | train | [[-0.82177734375,-0.82177734375,-0.82177734375,-0.82177734375,-0.82177734375,-0.82177734375,-0.82177(...TRUNCATED) | [50258,50364,50,569,20546,308,2394,1667,2265,72,2478,10895,275,65,25729,1667,369,5581,78,308,2394,16(...TRUNCATED) |
lin_100034 | lin | validation | [[-0.455810546875,-0.455810546875,-0.455810546875,-0.455810546875,-0.455810546875,-0.455810546875,-0(...TRUNCATED) | [50258,50364,37,6738,8835,17049,1714,64,11,8835,17049,308,2394,8588,13704,18603,9384,3016,11,8588,13(...TRUNCATED) |
lin_100038 | lin | train | [[-0.6650390625,-0.6650390625,-0.6650390625,-0.6650390625,-0.6650390625,-0.6650390625,-0.6650390625,(...TRUNCATED) | [50258,50364,45,304,897,9869,297,656,320,72,1714,64,20151,298,4037,37024,11,37024,5581,78,308,2394,8(...TRUNCATED) |
lin_100048 | lin | train | [[-1.5,-1.5,-1.5,-1.5,-1.5,-1.5,-1.0,-0.67578125,-0.75,-0.6123046875,-0.62060546875,-0.66552734375,-(...TRUNCATED) | [
50258,
50364,
33407,
8260,
72,
2478,
417,
453,
4711,
13,
50257
] |
End of preview. Expand in Data Studio
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]] | 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
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:
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.
- Downloads last month
- 46
Models trained or fine-tuned on mmwanje/waxal-features-v1