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 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 from Språkbanken. The subset contains 40M words from the social media genre between 2010 and 2015.

How to use

Example of transcribing 1% of the Common Voice test split, using GPU if available. The model expects 16kHz audio.

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

# Import model and processor
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 model is estimated. See this tutorial 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.