Wav2vec2 Large Xlsr 53 French Punctuation

🚀 Wav2Vec2 Large XLSR 53 French Punctuation Model

This model is fine - tuned for French speech recognition, capable of predicting text and punctuation, and has been evaluated on the Common Voice French dataset.

🚀 Quick Start

The following steps and code examples demonstrate how to evaluate this model on the Common Voice French test dataset.

💻 Usage Examples

Basic Usage

import re
import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import (
    Wav2Vec2ForCTC,
    Wav2Vec2Processor,
)

model_name = "Ilyes/wav2vec2-large-xlsr-53-french_punctuation"

model = Wav2Vec2ForCTC.from_pretrained(model_name).to('cuda')
processor = Wav2Vec2Processor.from_pretrained(model_name)

ds = load_dataset("common_voice", "fr", split="test")

chars_to_ignore_regex = '[\;\:\"\“\%\‘\”\�\‘\’\’\’\‘\…\·\ǃ\«\‹\»\›“\”\\ʿ\ʾ\„\∞\\|\;\:\*\—\–\─\―\_\/\:\ː\;\=\«\»\→]'
def normalize_text(text):
    text = text.lower().strip()
    text = re.sub('œ', 'oe', text)
    text = re.sub('æ', 'ae', text)
    text = re.sub("’|´|′|ʼ|‘|ʻ|`", "'", text)
    text = re.sub("'+ ", " ", text)
    text = re.sub(" '+", " ", text)
    text = re.sub("'$", " ", text)
    text = re.sub("' ", " ", text)
    text = re.sub("−|‐", "-", text)
    text = re.sub(" -", "", text)
    text = re.sub("- ", "", text)
    text = re.sub(chars_to_ignore_regex, '', text)
    return text

def map_to_array(batch):
    speech, _ = torchaudio.load(batch["path"])
    batch["speech"] = resampler.forward(speech.squeeze(0)).numpy()
    batch["sampling_rate"] = resampler.new_freq
    batch["sentence"] = normalize_text(batch["sentence"])
    return batch

ds = ds.map(map_to_array)

resampler = torchaudio.transforms.Resample(48_000, 16_000)
def map_to_pred(batch):
    features = processor(batch["speech"], sampling_rate=batch["sampling_rate"][0], padding=True, return_tensors="pt")
    input_values = features.input_values.to(device)
    attention_mask = features.attention_mask.to(device)
    with torch.no_grad():
        logits = model(input_values, attention_mask=attention_mask).logits
    pred_ids = torch.argmax(logits, dim=-1)
    batch["predicted"] = processor.batch_decode(pred_ids)
    batch["target"] = batch["sentence"]
    # remove duplicates
    batch["target"] = re.sub('\.+', '.', batch["target"])
    batch["target"] = re.sub('\?+', '?', batch["target"])
    batch["target"] = re.sub('!+', '!', batch["target"])
    batch["target"] = re.sub(',+', ',', batch["target"])
    return batch

result = ds.map(map_to_pred, batched=True, batch_size=16, remove_columns=list(ds.features.keys()))
wer = load_metric("wer")
print(wer.compute(predictions=result["predicted"], references=result["target"]))

📚 Documentation

Some Results

All the references and predictions of the test corpus are already available in this repository.

Overall Results

• Text + Punctuation: WER = 21.47%, CER = 7.21%
• Text (without punctuation): WER = 19.71%, CER = 6.91%

📄 License

This model is licensed under the Apache - 2.0 license.

📦 Model Information