KasuleTrevor/Swahili_100hrv2

wav2vec2-large-sw-cv-100hr-v2

This model is a fine-tuned version of facebook/wav2vec2-large on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.6293
• Model Preparation Time: 0.004
• Wer: 0.3558
• Cer: 0.1353

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0007
• train_batch_size: 32
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.6
• num_epochs: 120
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Model Preparation Time	Wer	Cer
3.5004	1.0	1040	0.7862	0.004	0.6822	0.1909
0.4704	2.0	2080	0.5245	0.004	0.4970	0.1335
0.3202	3.0	3120	0.4288	0.004	0.3956	0.1066
0.2551	4.0	4160	0.3800	0.004	0.3644	0.0962
0.2213	5.0	5200	0.3481	0.004	0.3358	0.0911
0.2005	6.0	6240	0.3241	0.004	0.3101	0.0839
0.1911	7.0	7280	0.3535	0.004	0.3283	0.0924
0.1854	8.0	8320	0.3161	0.004	0.3002	0.0843
0.1819	9.0	9360	0.3181	0.004	0.3233	0.0964
0.1788	10.0	10400	0.3099	0.004	0.3086	0.0832
0.1796	11.0	11440	0.3245	0.004	0.3011	0.0827
0.1805	12.0	12480	0.3068	0.004	0.2967	0.0845
0.1791	13.0	13520	0.3163	0.004	0.3076	0.0864
0.1812	14.0	14560	0.3336	0.004	0.3104	0.0867
0.1844	15.0	15600	0.3257	0.004	0.3117	0.0864
0.1884	16.0	16640	0.3440	0.004	0.3136	0.0886
0.1979	17.0	17680	0.3444	0.004	0.3172	0.0899
0.2058	18.0	18720	0.3286	0.004	0.3456	0.1092
0.2114	19.0	19760	0.3603	0.004	0.3358	0.0969
0.2229	20.0	20800	0.3658	0.004	0.3301	0.0954
0.2275	21.0	21840	0.3849	0.004	0.3729	0.1197
0.2351	22.0	22880	0.3753	0.004	0.3488	0.1011

Framework versions

• Transformers 4.44.0
• Pytorch 2.1.0+cu118
• Datasets 2.20.0
• Tokenizers 0.19.1