asr-africa/mms-300m_DigitalUmuganda_Afrivoice_Shona_50hr_v1

facebook/mms-300m

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

• Loss: 0.6190
• Wer: 0.3497
• Cer: 0.0791

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: 0.0003
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 150
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
22.4497	1.0	273	2.9247	1.0	1.0
8.283	2.0	546	0.5185	0.5978	0.1175
1.5316	3.0	819	0.2757	0.3547	0.0632
1.0241	4.0	1092	0.2358	0.3222	0.0570
0.9072	5.0	1365	0.2283	0.3081	0.0560
0.8161	6.0	1638	0.2201	0.2945	0.0527
0.7469	7.0	1911	0.2122	0.2818	0.0487
0.6981	8.0	2184	0.2191	0.2946	0.0519
0.6454	9.0	2457	0.2122	0.2892	0.0504
0.6094	10.0	2730	0.2206	0.2952	0.0498
0.5649	11.0	3003	0.2164	0.2732	0.0482
0.5197	12.0	3276	0.2143	0.2728	0.0479
0.4793	13.0	3549	0.2344	0.2766	0.0485
0.4368	14.0	3822	0.2266	0.2751	0.0486
0.3977	15.0	4095	0.2358	0.2669	0.0458
0.3667	16.0	4368	0.2447	0.2793	0.0474
0.3335	17.0	4641	0.2547	0.2692	0.0461
0.3044	18.0	4914	0.2566	0.2684	0.0463
0.2801	19.0	5187	0.2698	0.2737	0.0473
0.2514	20.0	5460	0.2967	0.2790	0.0480
0.2314	21.0	5733	0.2855	0.2708	0.0462
0.2092	22.0	6006	0.3074	0.2785	0.0476
0.1991	23.0	6279	0.2953	0.2681	0.0449
0.183	24.0	6552	0.3118	0.2712	0.0457
0.1706	25.0	6825	0.3108	0.2654	0.0454
0.1573	26.0	7098	0.3310	0.2730	0.0463
0.1538	27.0	7371	0.3142	0.2767	0.0463
0.1489	28.0	7644	0.3337	0.2666	0.0458
0.1404	29.0	7917	0.3176	0.2728	0.0462
0.1309	30.0	8190	0.3431	0.2707	0.0461
0.1246	31.0	8463	0.3368	0.2725	0.0458
0.1216	32.0	8736	0.3490	0.2648	0.0448
0.1205	33.0	9009	0.3436	0.2627	0.0447
0.1145	34.0	9282	0.3384	0.2693	0.0454
0.1117	35.0	9555	0.3500	0.2610	0.0443
0.1058	36.0	9828	0.3478	0.2608	0.0441
0.1039	37.0	10101	0.3434	0.2590	0.0438
0.1	38.0	10374	0.3507	0.2609	0.0438
0.0987	39.0	10647	0.3553	0.2610	0.0443
0.0977	40.0	10920	0.3691	0.2683	0.0451
0.0972	41.0	11193	0.3508	0.2657	0.0450
0.0946	42.0	11466	0.3445	0.2605	0.0445
0.0934	43.0	11739	0.3541	0.2628	0.0445
0.0896	44.0	12012	0.3424	0.2600	0.0437
0.0844	45.0	12285	0.3504	0.2599	0.0439
0.0847	46.0	12558	0.3630	0.2588	0.0433
0.0809	47.0	12831	0.3650	0.2565	0.0435
0.0812	48.0	13104	0.3632	0.2665	0.0446
0.0783	49.0	13377	0.3585	0.2629	0.0435
0.0739	50.0	13650	0.3722	0.2650	0.0435
0.0772	51.0	13923	0.3627	0.2607	0.0433
0.0746	52.0	14196	0.3712	0.2600	0.0439
0.0713	53.0	14469	0.3540	0.2664	0.0445
0.073	54.0	14742	0.3764	0.2571	0.0433
0.0712	55.0	15015	0.3625	0.2533	0.0428
0.0679	56.0	15288	0.3818	0.2603	0.0435
0.0701	57.0	15561	0.3921	0.2565	0.0434
0.0672	58.0	15834	0.3807	0.2584	0.0434
0.0681	59.0	16107	0.3787	0.2575	0.0438
0.063	60.0	16380	0.3818	0.2561	0.0428
0.062	61.0	16653	0.3919	0.2567	0.0429
0.0613	62.0	16926	0.4038	0.2598	0.0431
0.0617	63.0	17199	0.3885	0.2566	0.0431
0.0618	64.0	17472	0.3949	0.2557	0.0430
0.0572	65.0	17745	0.3839	0.2529	0.0418

Framework versions

• Transformers 4.47.0.dev0
• Pytorch 2.1.0+cu118
• Datasets 3.0.2
• Tokenizers 0.20.1