Hugging Face's logo

m3hrdadfi
/
wav2vec2-large-xlsr-icelandic

Automatic Speech Recognition
Transformers
PyTorch
JAX
Icelandic
wav2vec2
audio
speech
xlsr-fine-tuning-week
Eval Results
Inference Endpoints
Model card Files Community
1
wav2vec2-large-xlsr-icelandic / README.md
m3hrdadfi's picture
m3hrdadfi
Initial model
e432613
preview code
|
raw
history blame
6.72 kB
metadata
language: is
datasets:
  - malromur
tags:
  - audio
  - automatic-speech-recognition
  - speech
  - xlsr-fine-tuning-week
license: apache-2.0
widget:
  - label: Malromur sample 11
    src: >-
      https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/resolve/main/sample11.flac
  - label: Malromur sample 74
    src: >-
      https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/resolve/main/sample74.flac
model-index:
  - name: XLSR Wav2Vec2 Icelandic by Mehrdad Farahani
    results:
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Malromur is
          type: malromur
          args: lt
        metrics:
          - name: Test WER
            type: wer
            value: 12

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)[0]
    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={"remove_extra_space": True},
    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:

SOON

Evaluation

The model can be evaluated as follows on the 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

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)[0]
    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={"remove_extra_space": True},
    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: 12.00%

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.