MMS_cdli_ugandan_english_nonstandard_speech_finetune_v1

This model is a fine-tuned version of facebook/mms-300m on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 1.0944
• Model Preparation Time: 0.0089
• Wer: 0.4658
• Cer: 0.2311
• Normalized Wer: 0.4269
• Normalized Cer: 0.2206
• Semantic Error: 0.2363

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: 0.0001
• train_batch_size: 16
• eval_batch_size: 32
• seed: 42
• optimizer: Use OptimizerNames.ADAFACTOR and the args are: No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 40

Training results

Training Loss	Epoch	Step	Validation Loss	Model Preparation Time	Wer	Cer	Normalized Wer	Normalized Cer	Semantic Error
2.5623	3.0864	1000	2.4865	0.0089	1.0	0.7708	1.0	0.7672	0.9412
0.4631	6.1728	2000	1.2688	0.0089	0.5892	0.2926	0.5495	0.2802	0.3809
0.2057	9.2593	3000	1.2994	0.0089	0.5759	0.2897	0.5344	0.2788	0.3366
0.1254	12.3457	4000	1.4122	0.0089	0.5553	0.2881	0.5162	0.2771	0.3316
0.0863	15.4321	5000	1.6442	0.0089	0.5703	0.2934	0.5259	0.2803	0.2768
0.0424	18.5185	6000	1.8348	0.0089	0.5611	0.2893	0.5206	0.2783	0.3313
0.0189	21.6049	7000	2.1363	0.0089	0.5695	0.3017	0.5274	0.2886	0.3647

Framework versions

• Transformers 4.56.0
• Pytorch 2.8.0+cu129
• Datasets 2.18.0
• Tokenizers 0.22.0