wav2vec2-xls-r-1b-tevr

Overview

This folder contains a fully trained German speech recognition pipeline consisting of an acoustic model using the new wav2vec 2.0 XLS-R 1B TEVR architecture and a 5-gram KenLM language model. For an explanation of the TEVR enhancements and their motivation, please see our paper: TEVR: Improving Speech Recognition by Token Entropy Variance Reduction.

This pipeline scores a very competitive (as of June 2022) word error rate of 3.64% on CommonVoice German. The character error rate was 1.54%.

Citation

If you use this ASR pipeline for research, please cite:

@misc{https://doi.org/10.48550/arxiv.2206.12693,
  doi = {10.48550/ARXIV.2206.12693},
  url = {https://arxiv.org/abs/2206.12693},
  author = {Krabbenhöft, Hajo Nils and Barth, Erhardt},  
  keywords = {Computation and Language (cs.CL), Sound (cs.SD), Audio and Speech Processing (eess.AS), FOS: Computer and information sciences, FOS: Computer and information sciences, FOS: Electrical engineering, electronic engineering, information engineering, FOS: Electrical engineering, electronic engineering, information engineering, F.2.1; I.2.6; I.2.7},  
  title = {TEVR: Improving Speech Recognition by Token Entropy Variance Reduction},  
  publisher = {arXiv},  
  year = {2022}, 
  copyright = {Creative Commons Attribution 4.0 International}
}

TEVR Tokenizer Creation / Testing

See https://huggingface.co/fxtentacle/tevr-token-entropy-predictor-de for:

• our trained ByT5 model used to calculate the entropies in the paper
• a Jupyter Notebook to generate a TEVR Tokenizer from a text corpus
• a Jupyter Notebook to generate the illustration image in the paper

Evaluation

To evalue this pipeline yourself and/or on your own data, see the HF Eval Script.ipynb Jupyter Notebook or use the following python script:

!pip install --quiet --root-user-action=ignore --upgrade pip
!pip install --quiet --root-user-action=ignore "datasets>=1.18.3" "transformers==4.11.3" librosa jiwer huggingface_hub  
!pip install --quiet --root-user-action=ignore https://github.com/kpu/kenlm/archive/master.zip pyctcdecode
!pip install --quiet --root-user-action=ignore --upgrade transformers
!pip install --quiet --root-user-action=ignore torch_audiomentations audiomentations  

from datasets import load_dataset, Audio, load_metric
from transformers import AutoModelForCTC, Wav2Vec2ProcessorWithLM
import torchaudio.transforms as T
import torch
import unicodedata
import numpy as np
import re

# load testing dataset 
testing_dataset = load_dataset("common_voice", "de", split="test")

# replace invisible characters with space
allchars = list(set([c for t in testing_dataset['sentence'] for c in list(t)]))
map_to_space = [c for c in allchars if unicodedata.category(c)[0] in 'PSZ' and c not in 'ʻ-']
replacements = ''.maketrans(''.join(map_to_space), ''.join(' ' for i in range(len(map_to_space))), '\'ʻ')

def text_fix(text):
    # change ß to ss
    text = text.replace('ß','ss')
    # convert dash to space and remove double-space
    text = text.replace('-',' ').replace('  ',' ').replace('  ',' ')
    # make lowercase
    text = text.lower()
    # remap all invisible characters to space
    text = text.translate(replacements).strip()
    # for easier comparison to Zimmermeister, replace unrepresentable characters with ?
    text = re.sub("[âşěýňעảנźțãòàǔł̇æồאắîשðșęūāñë生בøúıśžçćńřğ]+","?",text)
    # remove multiple spaces (again)
    text = ' '.join([w for w in text.split(' ') if w != ''])
    return text

# load model
model = AutoModelForCTC.from_pretrained("fxtentacle/wav2vec2-xls-r-1b-tevr")
model.to('cuda')
# load processor
class HajoProcessor(Wav2Vec2ProcessorWithLM):
    @staticmethod
    def get_missing_alphabet_tokens(decoder, tokenizer):
        return []
processor = HajoProcessor.from_pretrained("fxtentacle/wav2vec2-xls-r-1b-tevr")

# this function will be called for each WAV file
def predict_single_audio(batch, image=False):    
    audio = batch['audio']['array']
    # resample, if needed
    if batch['audio']['sampling_rate'] != 16000:
        audio = T.Resample(orig_freq=batch['audio']['sampling_rate'], new_freq=16000)(torch.from_numpy(audio)).numpy()
    # normalize
    audio = (audio - audio.mean()) / np.sqrt(audio.var() + 1e-7)
    # ask HF processor to prepare audio for GPU eval
    input_values = processor(audio, return_tensors="pt", sampling_rate=16_000).input_values
    # call model on GPU
    with torch.no_grad():
        logits = model(input_values.to('cuda')).logits.cpu().numpy()[0]
    # ask HF processor to decode logits
    decoded = processor.decode(logits, beam_width=500)
    # return as dictionary
    return { 'groundtruth': text_fix(batch['sentence']), 'prediction': decoded.text }

# process all audio files
all_predictions = testing_dataset.map(predict_single_audio, remove_columns=testing_dataset.column_names)

# print results
print('WER', load_metric("wer").compute(predictions=all_predictions['prediction'], references=all_predictions['groundtruth'])*100.0, '%')
print('CER', load_metric("cer").compute(predictions=all_predictions['prediction'], references=all_predictions['groundtruth'])*100.0, '%')

WER 3.6433399042523233 %
CER 1.5398893560981173 %