Edit model card

wav2vec2-large-xls-r-1b-Swedish

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:

Without LM

  • Loss: 0.3370
  • Wer: 18.44
  • Cer: 5.75

With LM

  • Loss: 0.3370
  • Wer: 14.04
  • Cer: 4.86

Evaluation Commands

  1. To evaluate on mozilla-foundation/common_voice_8_0 with split test
python eval.py --model_id kingabzpro/wav2vec2-large-xls-r-1b-Swedish --dataset mozilla-foundation/common_voice_8_0 --config sv-SE --split test
  1. To evaluate on speech-recognition-community-v2/dev_data
python eval.py --model_id kingabzpro/wav2vec2-large-xls-r-1b-Swedish --dataset speech-recognition-community-v2/dev_data --config sv --split validation --chunk_length_s 5.0 --stride_length_s 1.0

Inference With LM

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-Swedish"
sample_iter = iter(load_dataset("mozilla-foundation/common_voice_8_0", "sv-SE", 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: 64
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 4
  • total_train_batch_size: 256
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 1000
  • num_epochs: 50
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer Cer
3.1562 11.11 500 0.4830 0.3729 0.1169
0.5655 22.22 1000 0.3553 0.2381 0.0743
0.3376 33.33 1500 0.3359 0.2179 0.0696
0.2419 44.44 2000 0.3232 0.1844 0.0575

Framework versions

  • Transformers 4.17.0.dev0
  • Pytorch 1.10.2+cu102
  • Datasets 1.18.2.dev0
  • Tokenizers 0.11.0
Downloads last month
109

Finetuned from

Dataset used to train kingabzpro/wav2vec2-large-xls-r-1b-Swedish

Collection including kingabzpro/wav2vec2-large-xls-r-1b-Swedish

Evaluation results