This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the common_voice dataset. It achieves the following results on the evaluation set:

Loss: 1.3599
Wer: 0.4236
Cer: 0.1768

Evaluation Commands

To evaluate on mozilla-foundation/common_voice_8_0 with split test

python eval.py --model_id kingabzpro/wav2vec2-large-xls-r-1b-Irish --dataset mozilla-foundation/common_voice_8_0 --config ga-IE --split test

Inference With LM

python

import torch
from datasets import load_dataset
from transformers import AutoModelForCTC, AutoProcessor
import torchaudio.functional as F
model_id = "kingabzpro/wav2vec2-large-xls-r-1b-Irish"
sample_iter = iter(load_dataset("mozilla-foundation/common_voice_8_0", "ga-IE", split="test", streaming=True, use_auth_token=True))
sample = next(sample_iter)
resampled_audio = F.resample(torch.tensor(sample["audio"]["array"]), 48_000, 16_000).numpy()
model = AutoModelForCTC.from_pretrained(model_id)
processor = AutoProcessor.from_pretrained(model_id)
input_values = processor(resampled_audio, return_tensors="pt").input_values
with torch.no_grad():
    logits = model(input_values).logits
transcription = processor.batch_decode(logits.numpy()).text

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 7.5e-05
train_batch_size: 32
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 128
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 200
num_epochs: 100
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
6.3955	12.48	100	2.9897	1.0	1.0
2.3811	24.97	200	1.2304	0.7140	0.3106
1.0476	37.48	300	1.0661	0.5597	0.2407
0.7014	49.97	400	1.1788	0.4799	0.1947
0.4409	62.48	500	1.2649	0.4658	0.1997
0.4839	74.97	600	1.3259	0.4450	0.1868
0.3643	87.48	700	1.3506	0.4312	0.1760
0.3468	99.97	800	1.3599	0.4236	0.1768

Framework versions

Transformers 4.17.0.dev0
Pytorch 1.10.2+cu102
Datasets 1.18.2.dev0
Tokenizers 0.11.0