wav2vec2-large-xlsr-persian

Wav2Vec2-Large-XLSR-53-Persian

Fine-tuned facebook/wav2vec2-large-xlsr-53 in Persian (Farsi) using Common Voice. When using this model, make sure that your speech input is sampled at 16kHz.

Usage

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

Requirements

# 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
!pip install hazm

Prediction

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

import numpy as np
import hazm
import re
import string

import IPython.display as ipd

_normalizer = hazm.Normalizer()
chars_to_ignore = [
    ",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "�",
    "#", "!", "؟", "?", "«", "»", "ء", "،", "(", ")", "؛", "'ٔ", "٬",'ٔ', ",", "?", 
    ".", "!", "-", ";", ":",'"',"“", "%", "‘", "”", "�", "–", "…", "_", "”", '“', '„'
]

# In case of farsi
chars_to_ignore = chars_to_ignore + list(string.ascii_lowercase + string.digits)

chars_to_mapping = {
    'ك': 'ک', 'دِ': 'د', 'بِ': 'ب', 'زِ': 'ز', 'ذِ': 'ذ', 'شِ': 'ش', 'سِ': 'س', 'ى': 'ی',
    'ي': 'ی', 'أ': 'ا', 'ؤ': 'و', "ے": "ی", "ۀ": "ه", "ﭘ": "پ", "ﮐ": "ک", "ﯽ": "ی",
    "ﺎ": "ا", "ﺑ": "ب", "ﺘ": "ت", "ﺧ": "خ", "ﺩ": "د", "ﺱ": "س", "ﻀ": "ض", "ﻌ": "ع",
    "ﻟ": "ل", "ﻡ": "م", "ﻢ": "م", "ﻪ": "ه", "ﻮ": "و", "ئ": "ی", 'ﺍ': "ا", 'ة': "ه",
    'ﯾ': "ی", 'ﯿ': "ی", 'ﺒ': "ب", 'ﺖ': "ت", 'ﺪ': "د", 'ﺮ': "ر", 'ﺴ': "س", 'ﺷ': "ش",
    'ﺸ': "ش", 'ﻋ': "ع", 'ﻤ': "م", 'ﻥ': "ن", 'ﻧ': "ن", 'ﻭ': "و", 'ﺭ': "ر", "ﮔ": "گ",
    "\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 = _normalizer.normalize(text)
    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-persian")
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-persian").to(device)

dataset = load_dataset("common_voice", "fa", 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), 20).tolist()
for i in max_items:
    reference, predicted =  result["sentence"][i], result["predicted"][i]
    print("reference:", reference)
    print("predicted:", predicted)
    print('---')

Output:

reference: اطلاعات مسری است
predicted: اطلاعات مسری است
---
reference: نه منظورم اینه که وقتی که ساکته چه کاریه خودمونه بندازیم زحمت
predicted: نه منظورم اینه که وقتی که ساکت چی کاریه خودمونو بندازیم زحمت
---
reference: من آب پرتقال می خورم لطفا
predicted: من آپ ارتغال می خورم لطفا
---
reference: وقت آن رسیده آنها را که قدم پیش میگذارند بزرگ بداریم
predicted: وقت آ رسیده آنها را که قدم پیش میگذارند بزرگ بداریم
---
reference: سیم باتری دارید
predicted: سیم باتری دارید
---
reference: این بهتره تا اینکه به بهونه درس و مشق هر روز بره خونه شون
predicted: این بهتره تا اینکه به بهمونه درسومش خرروز بره خونه اشون
---
reference: ژاکت تنگ است
predicted: ژاکت تنگ است
---
reference: آت و اشغال های خیابان
predicted: آت و اشغال های خیابان
---
reference: من به این روند اعتراض دارم
predicted: من به این لوند تراج دارم
---
reference: کرایه این مکان چند است
predicted: کرایه این مکان چند است
---
reference: ولی این فرصت این سهم جوانی اعطا نشده است
predicted: ولی این فرصت این سحم جوانی اتان نشده است
---
reference: متوجه فاجعهای محیطی میشوم
predicted: متوجه فاجایهای محیطی میشوم
---
reference: ترافیک شدیدیم بود و دیدن نور ماشینا و چراغا و لامپهای مراکز تجاری حس خوبی بهم میدادن
predicted: ترافیک شدید ی هم بودا دیدن نور ماشینا و چراغ لامپهای مراکز تجاری حس خولی بهم میدادن
---
reference: این مورد عمل ها مربوط به تخصص شما می شود
predicted: این مورد عملها مربوط به تخصص شما میشود
---
reference: انرژی خیلی کمی دارم
predicted: انرژی خیلی کمی دارم
---
reference: زیادی خوبی کردنم تهش داستانه
predicted: زیادی خوبی کردنم ترش داستانه
---
reference: بردهای که پادشاه شود
predicted: برده ای که پاده شاه شود
---
reference: یونسکو
predicted: یونسکو
---
reference: شما اخراج هستید
predicted: شما اخراج هستید
---
reference: من سفر کردن را دوست دارم
predicted: من سفر کردم را دوست دارم

Evaluation

The model can be evaluated as follows on the Persian (Farsi) 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 hazm
import re
import string

_normalizer = hazm.Normalizer()
chars_to_ignore = [
    ",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "�",
    "#", "!", "؟", "?", "«", "»", "ء", "،", "(", ")", "؛", "'ٔ", "٬",'ٔ', ",", "?", 
    ".", "!", "-", ";", ":",'"',"“", "%", "‘", "”", "�", "–", "…", "_", "”", '“', '„'
]

# In case of farsi
chars_to_ignore = chars_to_ignore + list(string.ascii_lowercase + string.digits)

chars_to_mapping = {
    'ك': 'ک', 'دِ': 'د', 'بِ': 'ب', 'زِ': 'ز', 'ذِ': 'ذ', 'شِ': 'ش', 'سِ': 'س', 'ى': 'ی',
    'ي': 'ی', 'أ': 'ا', 'ؤ': 'و', "ے": "ی", "ۀ": "ه", "ﭘ": "پ", "ﮐ": "ک", "ﯽ": "ی",
    "ﺎ": "ا", "ﺑ": "ب", "ﺘ": "ت", "ﺧ": "خ", "ﺩ": "د", "ﺱ": "س", "ﻀ": "ض", "ﻌ": "ع",
    "ﻟ": "ل", "ﻡ": "م", "ﻢ": "م", "ﻪ": "ه", "ﻮ": "و", "ئ": "ی", 'ﺍ': "ا", 'ة': "ه",
    'ﯾ': "ی", 'ﯿ': "ی", 'ﺒ': "ب", 'ﺖ': "ت", 'ﺪ': "د", 'ﺮ': "ر", 'ﺴ': "س", 'ﺷ': "ش",
    'ﺸ': "ش", 'ﻋ': "ع", 'ﻤ': "م", 'ﻥ': "ن", 'ﻧ': "ن", 'ﻭ': "و", 'ﺭ': "ر", "ﮔ": "گ",
    "\\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 = _normalizer.normalize(text)
    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-persian")
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-persian").to(device)

dataset = load_dataset("common_voice", "fa", 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: 32.20%

Training

The Common Voice train, validation datasets were used for training. The script used for training can be found here