README.md

---
language: sv
metrics:
- wer
tags:
- audio
- automatic-speech-recognition
- speech
- hf-asr-leaderboard
- sv
license: cc0-1.0
datasets:
- common_voice
- NST_Swedish_ASR_Database
- P4
- The_Swedish_Culturomics_Gigaword_Corpus
model-index:
- name: Wav2vec 2.0 large VoxRex Swedish (C) with 4-gram
  results:
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: Common Voice 6.1
      type: common_voice
      args: sv-SE
    metrics:
    - name: Test WER
      type: wer
      value: 6.4723
---

# KBLab's wav2vec 2.0 large VoxRex Swedish (C) with 4-gram model
Training of the acoustic model is the work of KBLab. See [VoxRex-C](https://huggingface.co/KBLab/wav2vec2-large-voxrex-swedish) for more details. This repo extends the acoustic model with a social media 4-gram language model for boosted performance.

## Model description
VoxRex-C is extended with a 4-gram language model estimated from a subset extracted from [The Swedish Culturomics Gigaword Corpus](https://spraakbanken.gu.se/resurser/gigaword) from Språkbanken. The subset contains 40M words from the social media genre between 2010 and 2015.

## How to use
#### Simple usage example with pipeline
```python
import torch
from transformers import pipeline

# Load the model. Using GPU if available
model_name = 'viktor-enzell/wav2vec2-large-voxrex-swedish-4gram'
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
pipe = pipeline(model=model_name).to(device)

# Run inference on an audio file
output = pipe('path/to/audio.mp3')['text']
```

#### More verbose usage example with audio pre-processing
Example of transcribing 1% of the Common Voice test split. The model expects 16kHz audio, so audio with another sampling rate is resampled to 16kHz.

```python
from transformers import Wav2Vec2ForCTC, Wav2Vec2ProcessorWithLM
from datasets import load_dataset
import torch
import torchaudio.functional as F

# Import model and processor. Using GPU if available
model_name = 'viktor-enzell/wav2vec2-large-voxrex-swedish-4gram'
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = Wav2Vec2ForCTC.from_pretrained(model_name).to(device);
processor = Wav2Vec2ProcessorWithLM.from_pretrained(model_name)

# Import and process speech data 
common_voice = load_dataset('common_voice', 'sv-SE', split='test[:1%]')

def speech_file_to_array(sample):
    # Convert speech file to array and downsample to 16 kHz
    sampling_rate = sample['audio']['sampling_rate']
    sample['speech'] = F.resample(torch.tensor(sample['audio']['array']), sampling_rate, 16_000)
    return sample

common_voice = common_voice.map(speech_file_to_array)

# Run inference
inputs = processor(common_voice['speech'], sampling_rate=16_000, return_tensors='pt', padding=True).to(device)

with torch.no_grad():
    logits = model(**inputs).logits

transcripts = processor.batch_decode(logits.cpu().numpy()).text
```

## Training procedure
Text data for the n-gram model is pre-processed by removing characters not part of the wav2vec 2.0 vocabulary and uppercasing all characters. After pre-processing and storing each text sample on a new line in a text file, a [KenLM](https://github.com/kpu/kenlm) model is estimated. See [this tutorial](https://huggingface.co/blog/wav2vec2-with-ngram) for more details.

## Evaluation results
The model was evaluated on the full Common Voice test set version 6.1. VoxRex-C achieved a WER of 9.03% without the language model and 6.47% with the language model.