README.md

metadata

language:
  - sv-SE
license: apache-2.0
tags:
  - automatic-speech-recognition
  - robust-speech-event
datasets:
  - mozilla-foundation/common_voice_8_0
metrics:
  - wer
  - cer
model-index:
  - name: wav2vec2-large-xls-r-1b-Swedish
    results:
      - task:
          type: automatic-speech-recognition
          name: Speech Recognition
        dataset:
          type: mozilla-foundation/common_voice_8_0
          name: Common Voice sv-SE
          args: sv-SE
        metrics:
          - type: wer
            value: 14.04
            name: Test WER With LM
            args:
              - learning_rate: 0.000075
              - 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: 1000
              - num_epochs: 50
              - mixed_precision_training: Native AMP
          - type: cer
            value: 4.86
            name: Test CER  With LM
            args:
              - learning_rate: 0.000075
              - 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: 1000
              - num_epochs: 50
              - mixed_precision_training: Native AMP

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

2. 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