Wav2vec2 Large Xlsr Estonian

🚀 Wav2Vec2-Large-XLSR-53-Estonian

This project fine-tunes the facebook/wav2vec2-large-xlsr-53 model in Estonian using Common Voice. Ensure your speech input is sampled at 16kHz when using this model.

🚀 Quick Start

The model can be used directly (without a language model) as follows:

💻 Usage Examples

Basic Usage

# requirement packages
!pip install git+https://github.com/huggingface/datasets.git
!pip install git+https://github.com/huggingface/transformers.git
!pip install torchaudio
!pip install librosa
!pip install jiwer

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

import numpy as np
import re
import string

import IPython.display as ipd

chars_to_ignore = [
    ",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "�",
    "#", "!", "?", "«", "»", "(", ")", "؛", ",", "?", ".", "!", "-", ";", ":", '"', 
    "“", "%", "‘", "�", "–", "…", "_", "”", '“', '„'
]
chars_to_mapping = {
"\u200c": " ", "\u200d": " ", "\u200e": " ", "\u200f": " ", "\ufeff": " ",
}

def multiple_replace(text, chars_to_mapping):
    pattern = "|".join(map(re.escape, chars_to_mapping.keys()))
    return re.sub(pattern, lambda m: chars_to_mapping[m.group()], str(text))

def remove_special_characters(text, chars_to_ignore_regex):
    text = re.sub(chars_to_ignore_regex, '', text).lower() + " "
    return text

def normalizer(batch, chars_to_ignore, chars_to_mapping):
    chars_to_ignore_regex = f"""[{"".join(chars_to_ignore)}]"""
    text = batch["sentence"].lower().strip()
    
    text = text.replace("\u0307", " ").strip()
    text = multiple_replace(text, chars_to_mapping)
    text = remove_special_characters(text, chars_to_ignore_regex)

    batch["sentence"] = text
    return batch


def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    speech_array = speech_array.squeeze().numpy()
    speech_array = librosa.resample(np.asarray(speech_array), sampling_rate, 16_000)

    batch["speech"] = speech_array
    return batch


def predict(batch):
    features = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

    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)[0]
    return batch


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
processor = Wav2Vec2Processor.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-estonian")
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-estonian").to(device)

dataset = load_dataset("common_voice", "et", split="test[:1%]")
dataset = dataset.map(
    normalizer, 
    fn_kwargs={"chars_to_ignore": chars_to_ignore, "chars_to_mapping": chars_to_mapping},
    remove_columns=list(set(dataset.column_names) - set(['sentence', 'path']))
)

dataset = dataset.map(speech_file_to_array_fn)
result = dataset.map(predict)

max_items = np.random.randint(0, len(result), 10).tolist()
for i in max_items:
    reference, predicted =  result["sentence"][i], result["predicted"][i]
    print("reference:", reference)
    print("predicted:", predicted)
    print('---')

Output:

reference: õhulossid lagunevad ning ees ootab maapind 
predicted: õhulassid lagunevad ning ees ootab maapind
---
reference: milliseks kiievisse pääsemise nimel võistlev muusik soome muusikamaastiku hetkeseisu hindab ning kas ta ka ennast sellel tulevikus tegutsemas näeb kuuled videost 
predicted: milliseks gievisse pääsemise nimel võitlev muusiks soome muusikama aastiku hetke seisu hindab ning kas ta ennast selle tulevikus tegutsemast näeb kuulad videost
---
reference: näiteks kui pool seina on tehtud tekib tunne et tahaks tegelikult natuke teistsugust ja hakkame otsast peale 
predicted: näiteks kui pool seine on tehtud tekib tunnetahaks tegelikult matuka teistsugust jahappanna otsast peane
---
reference: neuroesteetilised katsed näitavad et just nägude vaatlemine aktiveerib inimese aju esteetilist keskust 
predicted: neuroaisteetiliselt katsed näitaval et just nägude vaatlemine aptiveerid inimese aju est eedilist keskust
---
reference: paljud inimesed kindlasti kadestavad teid kuid ei julge samamoodi vabalt võtta 
predicted: paljud inimesed kindlasti kadestavadteid kuid ei julge sama moodi vabalt võtta
---
reference: parem on otsida pileteid inkognito veebi kaudu 
predicted: parem on otsida pileteid ning kognitu veebikaudu
---
reference: ja vot siin ma jäin vaikseks 
predicted: ja vat siisma ja invaikseks
---
reference: mida sa iseendale juubeli puhul soovid 
predicted: mida saise endale jubeli puhul soovid
---
reference: kuumuse ja kõrge temperatuuri tõttu kuivas tühjadel karjamaadel rohi mis muutus kergesti süttivaks 
predicted: kuumuse ja kõrge temperatuuri tõttu kuivast ühjadal karjamaadel rohi mis muutus kergesti süttivaks
---
reference: ilmselt on inimesi kelle jaoks on see hea lahendus 
predicted: ilmselt on inimesi kelle jaoks on see hea lahendus
---

Evaluation

The model can be evaluated as follows on the Estonian test data of Common Voice.

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

import numpy as np
import re
import string


chars_to_ignore = [
    ",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "�",
    "#", "!", "?", "«", "»", "(", ")", "؛", ",", "?", ".", "!", "-", ";", ":", '"', 
    "“", "%", "‘", "�", "–", "…", "_", "”", '“', '„'
]
chars_to_mapping = {
    "\u200c": " ", "\u200d": " ", "\u200e": " ", "\u200f": " ", "\ufeff": " ",
}

def multiple_replace(text, chars_to_mapping):
    pattern = "|".join(map(re.escape, chars_to_mapping.keys()))
    return re.sub(pattern, lambda m: chars_to_mapping[m.group()], str(text))

def remove_special_characters(text, chars_to_ignore_regex):
    text = re.sub(chars_to_ignore_regex, '', text).lower() + " "
    return text

def normalizer(batch, chars_to_ignore, chars_to_mapping):
    chars_to_ignore_regex = f"""[{"".join(chars_to_ignore)}]"""
    text = batch["sentence"].lower().strip()
    
    text = text.replace("\u0307", " ").strip()
    text = multiple_replace(text, chars_to_mapping)
    text = remove_special_characters(text, chars_to_ignore_regex)

    batch["sentence"] = text
    return batch


def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    speech_array = speech_array.squeeze().numpy()
    speech_array = librosa.resample(np.asarray(speech_array), sampling_rate, 16_000)

    batch["speech"] = speech_array
    return batch


def predict(batch):
    features = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

    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)[0]
    return batch


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
processor = Wav2Vec2Processor.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-estonian")
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-estonian").to(device)

dataset = load_dataset("common_voice", "et", split="test")
dataset = dataset.map(
    normalizer, 
    fn_kwargs={"chars_to_ignore": chars_to_ignore, "chars_to_mapping": chars_to_mapping},
    remove_columns=list(set(dataset.column_names) - set(['sentence', 'path']))
)

dataset = dataset.map(speech_file_to_array_fn)
result = dataset.map(predict)

wer = load_metric("wer")

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

Test Result:

• WER: 33.93%

📚 Documentation

The Common Voice train, validation datasets were used for training.

You can see the training states here

The script used for training can be found here

📄 License

This project is licensed under the Apache-2.0 license.