Wav2Vec2-Large-XLSR-53-ncj/nah

Fine-tuned facebook/wav2vec2-large-xlsr-53 on Nahuatl specifically of the Nort of Puebla (ncj) using a derivate of SLR92, and some samples of es and de datasets from Common Voice.

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", "{lang_id}", split="test[:2%]") # TODO: publish nahuatl_slr92_by_sentence

processor = Wav2Vec2Processor.from_pretrained("tyoc213/wav2vec2-large-xlsr-nahuatl")
model = Wav2Vec2ForCTC.from_pretrained("tyoc213/wav2vec2-large-xlsr-nahuatl")

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 Nahuatl specifically of the Nort of Puebla (ncj) 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", "{lang_id}", split="test") # TODO: publish nahuatl_slr92_by_sentence
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("tyoc213/wav2vec2-large-xlsr-nahuatl")
model = Wav2Vec2ForCTC.from_pretrained("tyoc213/wav2vec2-large-xlsr-nahuatl")
model.to("cuda")

chars_to_ignore_regex = '[\,\?\.\!\-\;\"\“\%\‘\”\�\(\)\-]'
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):
	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"] = 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: 50.95 %

Training

A derivate of SLR92 to be published soon.And some samples of es and de datasets from Common Voice

The script used for training can be found less60wer.ipynb