You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

One more step before getting this model.

This model is open access and available only for non-commercial use, with an SIL International AI & NLP RAIL-M license further specifying rights and usage.

The SIL RAIL-M License specifies:

You can't use the model to deliberately produce nor share illegal or harmful outputs or content. Particularly, you cannot use the model use with the intent or effect of harming or enabling discrimination against Indigenous People.
SIL claims no rights on outputs you generate for non-commercial use, you are free to use them and are accountable for their use, which must not go against the provisions set in the license
You may re-distribute the weights and use the model non-commercially including as a service. If you do, please be aware you have to include the same use restrictions as the ones in the license and share a copy of the SIL International AI & NLP RAIL-M to all your users (please read the license entirely and carefully). Please read the full license here: https://huggingface.co/spaces/sil-ai/model-license

By clicking on "Access repository" below, you accept that your contact information (email address and username) can be shared with the model authors as well.

If you would like to ask about commercial uses of this model, please email us.

wav2vec2-bloom-speech-eng

Model description

Homepage: SIL AI
Point of Contact: SIL AI email
Source Data: Bloom Library

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

Loss: 0.4718
Wer: 0.2112
Cer: 0.0650

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

Intended uses & limitations

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 and evaluation data

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

Training procedure

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

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0003
train_batch_size: 16
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 2
total_train_batch_size: 32
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 250
num_epochs: 1000.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
No log	2.02	250	2.9279	1.0	1.0
3.9597	4.03	500	2.2926	1.0	0.7266
3.9597	6.05	750	0.9448	0.7191	0.2751
1.2474	8.06	1000	0.5930	0.4661	0.1546
1.2474	10.08	1250	0.5095	0.3526	0.1127
0.5348	12.1	1500	0.4672	0.2789	0.0970
0.5348	14.11	1750	0.4874	0.2450	0.0769
0.3373	16.13	2000	0.4634	0.2669	0.0880
0.3373	18.14	2250	0.3927	0.2450	0.0752
0.2539	20.16	2500	0.4324	0.2530	0.0843
0.2539	22.18	2750	0.4436	0.2171	0.0695
0.2104	24.19	3000	0.5285	0.2311	0.0773
0.2104	26.21	3250	0.4844	0.2231	0.0695
0.1817	28.22	3500	0.4718	0.2112	0.0650
0.1817	30.24	3750	0.4518	0.2251	0.0683
0.1615	32.26	4000	0.3992	0.2131	0.0654
0.1615	34.27	4250	0.5397	0.2291	0.0736

Framework versions

Transformers 4.21.0.dev0
Pytorch 1.9.0+cu111
Datasets 2.2.2
Tokenizers 0.12.1

Downloads last month: 0

Inference Examples

Automatic Speech Recognition

This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Evaluation results

Test WER on Bloom Speech eng
self-reported

21.120
Test CER on Bloom Speech eng
self-reported

6.500

View on Papers With Code