joaoalvarenga/wav2vec2-large-100k-voxpopuli-pt

Wav2Vec2-Large-100k-VoxPopuli-Portuguese

Fine-tuned facebook/wav2vec2-large-100k-voxpopuli on Portuguese using the Common Voice dataset.

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", "pt", split="test[:2%]") 

processor = Wav2Vec2Processor.from_pretrained("joorock12/wav2vec2-large-100k-voxpopuli-pt")
model = Wav2Vec2ForCTC.from_pretrained("joorock12/wav2vec2-large-100k-voxpopuli-pt")

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["speech"][:2], 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["sentence"][:2])

Evaluation

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

You need to install Enelvo, an open-source spell correction trained with Twitter user posts pip install enelvo

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

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

processor = Wav2Vec2Processor.from_pretrained("joorock12/wav2vec2-large-100k-voxpopuli-pt")
model = Wav2Vec2ForCTC.from_pretrained("joorock12/wav2vec2-large-100k-voxpopuli-pt")
model.to("cuda")

chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�]'
resampler = torchaudio.transforms.Resample(48_000, 16_000)
norm = normaliser.Normaliser()

# Preprocessing the datasets.
# We need to read the aduio 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 aduio 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"] = [norm.normalise(i) for i in 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 (wer): 19.735723%

Training

The Common Voice train, validation datasets were used for training.