indonesian-nlp_wav2vec2-indonesian-javanese-sundanese

---

language:

• id
• jv
• sun datasets:
• mozilla-foundation/common_voice_7_0
• openslr
• magic_data
• titml metrics:
• wer tags:
• audio
• automatic-speech-recognition
• hf-asr-leaderboard
• id
• jv
• robust-speech-event
• speech
• su license: apache-2.0 model-index:
• name: Wav2Vec2 Indonesian Javanese and Sundanese by Indonesian NLP results:
  • task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: Common Voice 6.1 type: common_voice args: id metrics:
    • name: Test WER type: wer value: 4.056
    • name: Test CER type: cer value: 1.472
  • task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: Common Voice 7 type: mozilla-foundation/common_voice_7_0 args: id metrics:
    • name: Test WER type: wer value: 4.492
    • name: Test CER type: cer value: 1.577
  • task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: Robust Speech Event - Dev Data type: speech-recognition-community-v2/dev_data args: id metrics:
    • name: Test WER type: wer value: 48.94
  • task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: Robust Speech Event - Test Data type: speech-recognition-community-v2/eval_data args: id metrics:
    • name: Test WER type: wer value: 68.95

---

Multilingual Speech Recognition for Indonesian Languages

This is the model built for the project Multilingual Speech Recognition for Indonesian Languages. It is a fine-tuned facebook/wav2vec2-large-xlsr-53 model on the Indonesian Common Voice dataset, High-quality TTS data for Javanese - SLR41, and High-quality TTS data for Sundanese - SLR44 datasets.

We also provide a live demo to test the model.

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

test_dataset = load_dataset("common_voice", "id", split="test[:2%]")

processor = Wav2Vec2Processor.from_pretrained("indonesian-nlp/wav2vec2-indonesian-javanese-sundanese")
model = Wav2Vec2ForCTC.from_pretrained("indonesian-nlp/wav2vec2-indonesian-javanese-sundanese")

resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = resampler(speech_array).squeeze().numpy()
    return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)
inputs = processor(test_dataset[:2]["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

with torch.no_grad():
    logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits

predicted_ids = torch.argmax(logits, dim=-1)

print("Prediction:", processor.batch_decode(predicted_ids))
print("Reference:", test_dataset[:2]["sentence"])

Evaluation

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

import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re

test_dataset = load_dataset("common_voice", "id", split="test")
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("indonesian-nlp/wav2vec2-indonesian-javanese-sundanese")
model = Wav2Vec2ForCTC.from_pretrained("indonesian-nlp/wav2vec2-indonesian-javanese-sundanese") 
model.to("cuda")

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

resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the audio files as arrays
def speech_file_to_array_fn(batch):
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = resampler(speech_array).squeeze().numpy()
    return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)

# Preprocessing the datasets.
# We need to read the audio files as arrays
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 = test_dataset.map(evaluate, batched=True, batch_size=8)

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

Test Result: 11.57 %

Training

The Common Voice train, validation, and ... datasets were used for training as well as ... and ... # TODO

The script used for training can be found here (will be available soon)