Modelpublic

CLEAR-Global/whisper-small-clearglobal-kanuri-asr-1.0.0

sourceHugging Facecc-by-sa-4.0updated 7mo agoView on Hugging Face

2likes13downloads

Model Card

whisper-small-clearglobal-kanuri-asr-1.0.0

This model is a fine-tuned version of openai/whisper-small on the TWB Voice 1.0 dataset.

It achieves the following results on the evaluation set:

—WER: 10.16%
—Cer: 3.72%

Training and evaluation data

This model was trained by colleagues from the Makerere University Centre for Artificial Intelligence and Data Science in collaboration with CLEAR Global. We gratefully acknowledge their expertise and partnership.

Model was trained and tested on the approved Kanuri subset of TWB Voice 1.0 dataset.

Train/dev/test portions correspond to the splits in this dataset version.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

—learning_rate: 5e-05
—trainbatchsize: 16
—evalbatchsize: 16
—seed: 42
—gradientaccumulationsteps: 2
—totaltrainbatch_size: 32
—optimizer: Use OptimizerNames.ADAMWTORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizerargs=No additional optimizer arguments
—lrschedulertype: linear
—lrschedulerwarmup_ratio: 0.08
—num_epochs: 50.0
—mixedprecisiontraining: Native AMP

Training results

Training Loss	Epoch	Step	Cer	Validation Loss	Wer
1.3499	1.0	520	0.3772	0.4815	0.8971
0.3923	2.0	1040	0.1944	0.3327	0.6041
0.2181	3.0	1560	0.1242	0.2628	0.4174
0.1286	4.0	2080	0.0937	0.2261	0.3440
0.0832	5.0	2600	0.0671	0.1934	0.2607
0.0512	6.0	3120	0.0599	0.1768	0.2275
0.0356	7.0	3640	0.0548	0.1630	0.1975
0.0278	8.0	4160	0.0514	0.1581	0.1847
0.0226	9.0	4680	0.0466	0.1572	0.1696
0.0175	10.0	5200	0.0420	0.1518	0.1541
0.0153	11.0	5720	0.0438	0.1557	0.1533
0.0125	12.0	6240	0.0407	0.1444	0.1437
0.0113	13.0	6760	0.0404	0.1469	0.1424
0.0098	14.0	7280	0.0414	0.1449	0.1442
0.0082	15.0	7800	0.0371	0.1401	0.1323
0.0078	16.0	8320	0.0406	0.1374	0.1374
0.0072	17.0	8840	0.0373	0.1408	0.1297
0.0059	18.0	9360	0.0370	0.1374	0.1277
0.0063	19.0	9880	0.0370	0.1347	0.1231
0.0057	20.0	10400	0.0349	0.1386	0.1185
0.0046	21.0	10920	0.0347	0.1346	0.1185
0.0043	22.0	11440	0.0359	0.1410	0.1218
0.0041	23.0	11960	0.0330	0.1296	0.1125
0.0029	24.0	12480	0.0330	0.1308	0.1110
0.0033	25.0	13000	0.0384	0.1364	0.1191
0.0036	26.0	13520	0.0318	0.1264	0.1073
0.0027	27.0	14040	0.0325	0.1264	0.1074
0.0016	28.0	14560	0.0322	0.1257	0.1046
0.0015	29.0	15080	0.0322	0.1257	0.1032
0.0018	30.0	15600	0.0303	0.1251	0.1016
0.0018	31.0	16120	0.0326	0.1332	0.1071
0.0024	32.0	16640	0.0319	0.1282	0.1073
0.0008	33.0	17160	0.0309	0.1256	0.1018
0.0007	34.0	17680	0.0297	0.1250	0.0970
0.0009	35.0	18200	0.1305	0.1057	0.0335
0.0012	36.0	18720	0.1312	0.0980	0.0299
0.0009	37.0	19240	0.1307	0.1004	0.0309
0.0005	38.0	19760	0.1263	0.0960	0.0293
0.0004	39.0	20280	0.1263	0.0933	0.0285
0.0002	40.0	20800	0.1273	0.0935	0.0283
0.0001	41.0	21320	0.1262	0.0916	0.0281
0.0001	42.0	21840	0.1267	0.0926	0.0287
0.0	43.0	22360	0.1271	0.0907	0.0277
0.0	44.0	22880	0.1275	0.0900	0.0274
0.0	45.0	23400	0.1279	0.0893	0.0273
0.0	46.0	23920	0.1282	0.0884	0.0267

Framework versions

—Transformers 4.53.1
—Pytorch 2.7.1+cu126
—Datasets 4.0.0
—Tokenizers 0.21.2