wav2vec2-bloom-speech-bam

logo for Bloom Library sil-ai logo

Model description

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the SIL-AI/bloom-speech - BAM (Bambara) dataset. It achieves the following results on the evaluation set:

Users should refer to the original model for tutorials on using a trained model for inference.

Users of this model must abide by the SIL RAIL-M License.

This model is created as a proof of concept and no guarantees are made regarding the performance of the model is specific situations.

Training, Validation, and Test datasets were generated from the same corpus, ensuring that no duplicate files were used.

Standard finetuning of XLS-R was used based on the examples in the Hugging Face Transformers Github

The following hyperparameters were used during training:

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
No log	41.67	250	0.8321	0.5770	0.1769
2.495	83.33	500	0.7474	0.4692	0.1418
2.495	125.0	750	0.8151	0.4159	0.1241
0.07	166.67	1000	0.8168	0.4645	0.1439
0.07	208.33	1250	0.8188	0.3981	0.1186
0.0361	250.0	1500	0.8344	0.4028	0.1202
0.0361	291.67	1750	0.9109	0.4301	0.1306
0.023	333.33	2000	0.9366	0.4088	0.1210
0.023	375.0	2250	0.9443	0.4088	0.1158
0.0187	416.67	2500	1.0506	0.4289	0.1316
0.0187	458.33	2750	1.0120	0.3780	0.1137
0.0159	500.0	3000	1.0376	0.3768	0.1139
0.0159	541.67	3250	1.0098	0.4088	0.1202
0.0145	583.33	3500	1.0302	0.3993	0.1132
0.0145	625.0	3750	1.0513	0.3910	0.1134
0.0116	666.67	4000	1.0325	0.3934	0.1150
0.0116	708.33	4250	1.0285	0.3898	0.1124
0.0109	750.0	4500	0.9712	0.3626	0.1134
0.0109	791.67	4750	1.0263	0.3614	0.1113
0.0089	833.33	5000	1.0592	0.3590	0.1093
0.0089	875.0	5250	1.0888	0.3614	0.1093
0.0082	916.67	5500	1.0896	0.3661	0.1108
0.0082	958.33	5750	1.0890	0.3626	0.1108