wav2vec 2.0 multilingual ( Finetued )
The base model pretrained on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz. Note that this model should be fine-tuned on a downstream task, like Automatic Speech Recognition. Check out this blog for more information.
Authors: Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli
Abstract This paper presents XLSR which learns cross-lingual speech representations by pretraining a single model from the raw waveform of speech in multiple languages. We build on wav2vec 2.0 which is trained by solving a contrastive task over masked latent speech representations and jointly learns a quantization of the latents shared across languages. The resulting model is fine-tuned on labeled data and experiments show that cross-lingual pretraining significantly outperforms monolingual pretraining. On the CommonVoice benchmark, XLSR shows a relative phoneme error rate reduction of 72% compared to the best known results. On BABEL, our approach improves word error rate by 16% relative compared to a comparable system. Our approach enables a single multilingual speech recognition model which is competitive to strong individual models. Analysis shows that the latent discrete speech representations are shared across languages with increased sharing for related languages. We hope to catalyze research in low-resource speech understanding by releasing XLSR-53, a large model pretrained in 53 languages.
The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/wav2vec#wav2vec-20.
Fine-tuned facebook/wav2vec2-large-xlsr-53 in Persian (Farsi) using Common Voice plus Our own created Dataset(1/3 of total dataset). When using this model, make sure that your speech input is sampled at 16kHz.
Evaluation: π‘οΈ
We have evaluated the model on private dataset with different type of audios (unfortunately the dataset for testing and validation is not publicly available but to see a sample of the dataset check this link) :
Name | test dataset (wer) |
---|---|
m3hrdadfi/wav2vec2-large-xlsr-persian-v3 | 0.56754 |
This New Model | 0.40815 |
Base Multilingual Model | 0.69746 |
- This Table show if we add more data we will have much better result
How to useβ
Use FineTuned Model
This model is finetuned on m3hrdadfi/wav2vec2-large-xlsr-persian-v3 , so the process for train or evaluation is same
# 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 parsivar
!pip install num2fawords
Normalizer
# Normalizer
!wget -O normalizer.py https://huggingface.co/m3hrdadfi/"wav2vec2-large-xlsr-persian-v3/raw/main/dictionary.py
!wget -O normalizer.py https://huggingface.co/m3hrdadfi/"wav2vec2-large-xlsr-persian-v3/raw/main/normalizer.py
If you are not sure your transcriptions are clean or not (having weird characters or any other alphabete chars ) use this code provided by m3hrdadfi/wav2vec2-large-xlsr-persian-v3
Cleaning (Fill the data part with your own data dir)
from normalizer import normalizer
def cleaning(text):
if not isinstance(text, str):
return None
return normalizer({"sentence": text}, return_dict=False)
# edit these parts with your own data directory
data_dir = "data"
test = pd.read_csv(f"{data_dir}/yourtest.tsv", sep=" ")
test["path"] = data_dir + "/clips/" + test["path"]
print(f"Step 0: {len(test)}")
test["status"] = test["path"].apply(lambda path: True if os.path.exists(path) else None)
test = test.dropna(subset=["path"])
test = test.drop("status", 1)
print(f"Step 1: {len(test)}")
test["sentence"] = test["sentence"].apply(lambda t: cleaning(t))
test = test.dropna(subset=["sentence"])
print(f"Step 2: {len(test)}")
test = test.reset_index(drop=True)
print(test.head())
test = test[["path", "sentence"]]
test.to_csv("/content/test.csv", sep=" ", encoding="utf-8", index=False)
Prediction
import numpy as np
import pandas as pd
import librosa
import torch
import torchaudio
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
from datasets import load_dataset, load_metric
import IPython.display as ipd
model_name_or_path = "masoudmzb/wav2vec2-xlsr-multilingual-53-fa"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(model_name_or_path, device)
processor = Wav2Vec2Processor.from_pretrained(model_name_or_path)
model = Wav2Vec2ForCTC.from_pretrained(model_name_or_path).to(device)
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, processor.feature_extractor.sampling_rate)
batch["speech"] = speech_array
return batch
def predict(batch):
features = processor(
batch["speech"],
sampling_rate=processor.feature_extractor.sampling_rate,
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)
return batch
# edit these parts with your own data directory
dataset = load_dataset("csv", data_files={"test": "/path_to/your_test.csv"}, delimiter=" ")["test"]
dataset = dataset.map(speech_file_to_array_fn)
result = dataset.map(predict, batched=True, batch_size=4)
WER Score
wer = load_metric("wer")
print("WER: {:.2f}".format(100 * wer.compute(predictions=result["predicted"], references=result["sentence"])))
Output
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('---')
training details: π
One model was trained on Persian Mozilla dataset before So we Decided to continue from that one. Model is warm started from mehrdadfa
βs checkpoint
- For more details, you can take a look at config.json at the model card in π€ Model Hub
- The model trained 84000 steps, equal to 12.42 Epochs.
- The base model to finetune was https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-persian-v3/tree/main
Fine Tuning Recommendations: π€
For fine tuning you can check the link below. but be aware some Tips. you may need gradient_accumulation because you need more batch size. the are many hyperparameters make sure you set them properly :
- learning_rate
- attention_dropout
- hidden_dropout
- feat_proj_dropout
- mask_time_prob
- layer_drop
Fine Tuning Examples π·ββοΈπ·ββοΈ
Dataset | Fine Tuning Example |
---|---|
Fine Tune on Mozilla Turkish Dataset | |
Sample Code for Other Dataset And other Language | github_link |
Contact us: π€
If you have a technical question regarding the model, pretraining, code or publication, please create an issue in the repository. This is the fastest way to reach us.
Citation: β©οΈ
we didn't publish any papers on the work. However, if you did, please cite us properly with an entry like one below.
@misc{wav2vec2-xlsr-multilingual-53-fa,
author = {Paparnchi, Seyyed Mohammad Masoud},
title = {wav2vec2-xlsr-multilingual-53-fa},
year = 2021,
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/Hamtech-ai/wav2vec2-fa}},
}
- Downloads last month
- 145