othrif/wav2vec2-large-xlsr-moroccan

Wav2Vec2-Large-XLSR-53-Moroccan

Fine-tuned facebook/wav2vec2-large-xlsr-53 on MGB5 Moroccan Arabic kindly provided by ELDA and ArabicSpeech.

In order to have access to MGB5, please request it from ELDA.

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:

import re
import torch
import librosa
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import soundfile as sf


dataset = load_dataset("ma_speech_corpus", split="test")

processor = Wav2Vec2Processor.from_pretrained("othrif/wav2vec2-large-xlsr-moroccan")
model = Wav2Vec2ForCTC.from_pretrained("othrif/wav2vec2-large-xlsr-moroccan")
model.to("cuda")


chars_to_ignore_regex = '[\\,\\?\\.\\!\\-\\;\\:\\"\\“\\'\\�]'

def remove_special_characters(batch):
    batch["text"] = re.sub(chars_to_ignore_regex, "", batch["sentence"]).lower() + " "
    return batch


dataset = dataset.map(remove_special_characters)
dataset = dataset.select(range(10))

def speech_file_to_array_fn(batch):
    start, stop = batch['segment'].split('_')
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    speech_array, sampling_rate = sf.read(batch["path"], start=int(float(start) * sampling_rate),
                                          stop=int(float(stop) * sampling_rate))
    batch["speech"] = librosa.resample(speech_array, sampling_rate, 16_000)
    batch["sampling_rate"] = 16_000
    batch["target_text"] = batch["text"]
    return batch


dataset = dataset.map(
    speech_file_to_array_fn
)

def predict(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["predicted"] = processor.batch_decode(pred_ids)
    return batch

dataset = dataset.map(predict, batched=True, batch_size=32)

for reference, predicted in zip(dataset["sentence"], dataset["predicted"]):
    print("reference:", reference)
    print("predicted:", predicted)
    print("--")

Here's the output:

reference: عشرين ألفريال الوحده وشي خمسميه دريال

predicted: عشرين علف ريا لوحده وشي خمسميات ريال
--
reference: واحد جوج تلاتة ربعه خمسة ستة

predicted: غيحك تويش تتبة نتاست
--
reference: هي هاديك غتجينا تقريبا ميه وسته وعشرين ألف ريال

predicted: ياض كتجينا تقريبه ميه أو ستي و عشيناأفرين
--
reference: ###والصرف ليبقا نجيب بيه الصالون فلهوندا... أهاه نديروها علاش لا؟...

predicted: أواصرف ليبقا نجيب يه اصالون فالهندا أه نديروها علاش لا
--
reference: ###صافي مشات... أنا أختي معندي مندير بهاد صداع الراس...

predicted: صافي مشات أنا خصي معندي مندير بهاد داع راسك
  ف
--
reference: خلصو ليا غير لكريدي ديالي وديرو ليعجبكوم

predicted: خلصو ليا غير لكريدي ديالي أوديرو لي عجبكوم
--
reference: أنا نتكلف يلاه لقى شي حاجه نشغل بيها راسي

predicted: أنا نتكلف يالله لقا شي حاجه نشغل بيها راسي

Evaluation

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

import re
import torch
import librosa
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import soundfile as sf

eval_dataset = load_dataset("ma_speech_corpus", split="test")
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("othrif/wav2vec2-large-xlsr-moroccan")
model = Wav2Vec2ForCTC.from_pretrained("othrif/wav2vec2-large-xlsr-moroccan")
model.to("cuda")

chars_to_ignore_regex = '[\\,\\?\\.\\!\\-\\;\\:\\"\\“\\'\\�]'

def remove_special_characters(batch):
    batch["text"] = re.sub(chars_to_ignore_regex, "", batch["sentence"]).lower() + " "
    return batch


eval_dataset = eval_dataset.map(remove_special_characters, remove_columns=["sentence"])
#eval_dataset = eval_dataset.select(range(100))

def speech_file_to_array_fn(batch):
    start, stop = batch['segment'].split('_')
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    speech_array, sampling_rate = sf.read(batch["path"], start=int(float(start) * sampling_rate),
                                          stop=int(float(stop) * sampling_rate))
    batch["speech"] = librosa.resample(speech_array, sampling_rate, 16_000)
    batch["sampling_rate"] = 16_000
    batch["target_text"] = batch["text"]
    return batch


eval_dataset = eval_dataset.map(
    speech_file_to_array_fn,
    remove_columns=eval_dataset.column_names
)

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 = eval_dataset.map(evaluate, batched=True, batch_size=32)

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

Test Result: 66.45

Training

The MGB5 train, validation datasets were used for training.

The script used for training can be found here