m3hrdadfi/wav2vec2-large-xlsr-icelandic

Wav2Vec2-Large-XLSR-53-Icelandic

Fine-tuned facebook/wav2vec2-large-xlsr-53 in Icelandic using Malromur. 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 num2words

Normalizer

# num2word packages
# Original source: https://github.com/savoirfairelinux/num2words
!mkdir -p ./num2words
!wget -O num2words/__init__.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/num2words/__init__.py
!wget -O num2words/base.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/num2words/base.py
!wget -O num2words/compat.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/num2words/compat.py
!wget -O num2words/currency.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/num2words/currency.py
!wget -O num2words/lang_EU.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/num2words/lang_EU.py
!wget -O num2words/lang_IS.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/num2words/lang_IS.py
!wget -O num2words/utils.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/num2words/utils.py

# Malromur_test selected based on gender and age
!wget -O malromur_test.csv https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/malromur_test.csv

# Normalizer
!wget -O normalizer.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/normalizer.py

Prediction

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

from normalizer import Normalizer

normalizer = Normalizer(lang="is")


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)
    return batch


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

dataset = load_dataset("csv", data_files={"test": "./malromur_test.csv"})["test"]
dataset = dataset.map(
    normalizer, 
    fn_kwargs={"do_lastspace_removing": True, "text_key_name": "cleaned_sentence"},
    remove_columns=list(set(dataset.column_names) - set(['cleaned_sentence', 'path']))
)

dataset = dataset.map(speech_file_to_array_fn)
result = dataset.map(predict, batched=True, batch_size=8)

max_items = np.random.randint(0, len(result), 20).tolist()
for i in max_items:
    reference, predicted =  result["cleaned_sentence"][i], result["predicted"][i]
    print("reference:", reference)
    print("predicted:", predicted)
    print('---')

Output: ```text reference: eða eitthvað annað dýr predicted: eða eitthvað annað dýr

reference: oddgerður predicted: oddgerður

reference: eiðný predicted: eiðný

reference: löndum predicted: löndum

reference: tileinkaði bróður sínum markið predicted: tileinkaði bróður sínum markið

reference: þetta er svo mikill hégómi predicted: þetta er svo mikill hégómi

reference: timarit is predicted: timarit is

reference: stefna strax upp aftur predicted: stefna strax upp aftur

reference: brekkuflöt predicted: brekkuflöt

reference: áætlunarferð frestað vegna veðurs predicted: áætluna ferð frestað vegna veðurs

reference: sagði af sér vegna kláms predicted: sagði af sér vegni kláms

reference: grímúlfur predicted: grímúlgur

reference: lýsti sig saklausan predicted: lýsti sig saklausan

reference: belgingur is predicted: belgingur is

reference: sambía predicted: sambía

reference: geirastöðum predicted: geirastöðum

reference: varð tvisvar fyrir eigin bíl predicted: var tvisvar fyrir eigin bíl

reference: reykjavöllum predicted: reykjavöllum

reference: miklir menn eru þeir þremenningar predicted: miklir menn eru þeir þremenningar

reference: handverkoghonnun is predicted: handverkoghonnun is

## Evaluation

The model can be evaluated as follows on the test data of Malromur.

```python
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

from normalizer import Normalizer

normalizer = Normalizer(lang="is")


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)
    return batch


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

dataset = load_dataset("csv", data_files={"test": "./malromur_test.csv"})["test"]
dataset = dataset.map(
    normalizer, 
    fn_kwargs={"do_lastspace_removing": True, "text_key_name": "cleaned_sentence"},
    remove_columns=list(set(dataset.column_names) - set(['cleaned_sentence', 'path']))
)

dataset = dataset.map(speech_file_to_array_fn)
result = dataset.map(predict, batched=True, batch_size=8)

wer = load_metric("wer")

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

Test Result:

• WER: 09.21%

Training & Report

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

Questions?

Post a Github issue on the Wav2Vec repo.