Wav2vec2 Large Xlsr 53 German

🚀 Wav2Vec2-Large-XLSR-53-German

This model is fine-tuned from facebook/wav2vec2-large-xlsr-53 on German using the Common Voice dataset, which is useful for automatic speech recognition tasks.

🚀 Quick Start

When using this model, make sure that your speech input is sampled at 16kHz.

✨ Features

• Fine - tuned on German language data from the Common Voice dataset.
• Can be used for automatic speech recognition tasks without a language model.

📦 Installation

No specific installation steps are provided in the original document, so this section is skipped.

💻 Usage Examples

Basic Usage

import torch
import torchaudio
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

test_dataset = load_dataset("common_voice", "de", split="test[:2%]") 

processor = Wav2Vec2Processor.from_pretrained("marcel/wav2vec2-large-xlsr-53-german")
model = Wav2Vec2ForCTC.from_pretrained("marcel/wav2vec2-large-xlsr-53-german") 

resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
	speech_array, sampling_rate = torchaudio.load(batch["path"])
	batch["speech"] = resampler(speech_array).squeeze().numpy()
	return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)
inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)

with torch.no_grad():
	logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits

predicted_ids = torch.argmax(logits, dim=-1)

print("Prediction:", processor.batch_decode(predicted_ids))
print("Reference:", test_dataset["sentence"][:2])

Advanced Usage

The model can also be evaluated in different ways. The following is an example of evaluating the model on the entire test data of Common Voice:

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

test_dataset = load_dataset("common_voice", "de", split="test")
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("marcel/wav2vec2-large-xlsr-53-german")
model = Wav2Vec2ForCTC.from_pretrained("marcel/wav2vec2-large-xlsr-53-german") 
model.to("cuda")

chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\”\�\カ\æ\無\ན\カ\臣\ѹ\…\«\»\ð\ı\„\幺\א\ב\比\ш\ע\)\ứ\в\œ\ч\+\—\ш\‚\נ\м\ń\乡\$\=\ש\ф\支\(\°\и\к\̇]'
substitutions = {
    'e' : '[\ə\é\ě\ę\ê\ế\ế\ë\ė\е]',
    'o' : '[\ō\ô\ô\ó\ò\ø\ọ\ŏ\õ\ő\о]',
    'a' : '[\á\ā\ā\ă\ã\å\â\à\ą\а]',
    'c' : '[\č\ć\ç\с]',
    'l' : '[\ł]',
    'u' : '[\ú\ū\ứ\ů]',
    'und' : '[\&]',
    'r' : '[\ř]',
    'y' : '[\ý]',
    's' : '[\ś\š\ș\ş]',
    'i' : '[\ī\ǐ\í\ï\î\ï]',
    'z' : '[\ź\ž\ź\ż]',
    'n' : '[\ñ\ń\ņ]',
    'g' : '[\ğ]',
    'ss' : '[\ß]',
    't' : '[\ț\ť]',
    'd' : '[\ď\đ]',
    "'": '[\ʿ\་\’\`\´\ʻ\`\‘]',
    'p': '\р'
}
resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
    for x in substitutions:
        batch["sentence"] = re.sub(substitutions[x], x, batch["sentence"])
        speech_array, sampling_rate = torchaudio.load(batch["path"])
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = resampler(speech_array).squeeze().numpy()
    return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def evaluate(batch):
    inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

    with torch.no_grad():
        logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits

    pred_ids = torch.argmax(logits, dim=-1)
    batch["pred_strings"] = processor.batch_decode(pred_ids)
    return batch

result = test_dataset.map(evaluate, batched=True, batch_size=8)

print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))

The model can also be evaluated with in 10% chunks which needs less ressources (to be tested).

import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re
import jiwer
lang_id = "de"

processor = Wav2Vec2Processor.from_pretrained("marcel/wav2vec2-large-xlsr-53-german")
model = Wav2Vec2ForCTC.from_pretrained("marcel/wav2vec2-large-xlsr-53-german") 
model.to("cuda")

chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\”\�\カ\æ\無\ན\カ\臣\ѹ\…\«\»\ð\ı\„\幺\א\ב\比\ш\ע\)\ứ\в\œ\ч\+\—\ш\‚\נ\м\ń\乡\$\=\ש\ф\支\(\°\и\к\̇]'
substitutions = {
    'e' : '[\ə\é\ě\ę\ê\ế\ế\ë\ė\е]',
    'o' : '[\ō\ô\ô\ó\ò\ø\ọ\ŏ\õ\ő\о]',
    'a' : '[\á\ā\ā\ă\ã\å\â\à\ą\а]',
    'c' : '[\č\ć\ç\с]',
    'l' : '[\ł]',
    'u' : '[\ú\ū\ứ\ů]',
    'und' : '[\&]',
    'r' : '[\ř]',
    'y' : '[\ý]',
    's' : '[\ś\š\ș\ş]',
    'i' : '[\ī\ǐ\í\ï\î\ï]',
    'z' : '[\ź\ž\ź\ż]',
    'n' : '[\ñ\ń\ņ]',
    'g' : '[\ğ]',
    'ss' : '[\ß]',
    't' : '[\ț\ť]',
    'd' : '[\ď\đ]',
    "'": '[\ʿ\་\’\`\´\ʻ\`\‘]',
    'p': '\р'
}
resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
    for x in substitutions:
        batch["sentence"] = re.sub(substitutions[x], x, batch["sentence"])
        speech_array, sampling_rate = torchaudio.load(batch["path"])
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = resampler(speech_array).squeeze().numpy()
    return batch

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def evaluate(batch):
    inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

    with torch.no_grad():
        logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits

    pred_ids = torch.argmax(logits, dim=-1)
    batch["pred_strings"] = processor.batch_decode(pred_ids)
    return batch

H, S, D, I = 0, 0, 0, 0
for i in range(10):
    print("test["+str(10*i)+"%:"+str(10*(i+1))+"%]")
    test_dataset = load_dataset("common_voice", "de", split="test["+str(10*i)+"%:"+str(10*(i+1))+"%]")
    test_dataset = test_dataset.map(speech_file_to_array_fn)
    result = test_dataset.map(evaluate, batched=True, batch_size=8)
    predictions = result["pred_strings"]
    targets = result["sentence"]
    chunk_metrics = jiwer.compute_measures(targets, predictions)
    H = H + chunk_metrics["hits"]
    S = S + chunk_metrics["substitutions"]
    D = D + chunk_metrics["deletions"]
    I = I + chunk_metrics["insertions"]
WER = float(S + D + I) / float(H + S + D)
print("WER: {:2f}".format(WER*100))

Test Result: 15.80 %

📚 Documentation

Training Details

The first 50% of the Common Voice train, and 12% of the validation datasets were used for training (30 epochs on first 12% and 3 epochs on the remainder).

🔧 Technical Details

No specific technical details are provided in the original document, so this section is skipped.

📄 License

This project is licensed under the Apache - 2.0 license.