Ontocord/mt5-fix-asr-vietnamese

Fine-tuned mt5 to correct output of an ASR model trained on facebook/wav2vec2-large-xlsr-53 which was trained on Vietnamese using the Common Voice, and FOSD.

Usage

The model can be used directly by submitting vietnamese asr text, but is is best to use with the ontocord/wav2vec2-large-xlsr-vietnamese model.

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

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
test_dataset = load_dataset("common_voice", "vi", split="test[:2%]") 
processor = Wav2Vec2Processor.from_pretrained("ontocord/wav2vec2-large-xlsr-53-vietnamese") 
model = Wav2Vec2ForCTC.from_pretrained("ontocord/wav2vec2-large-xlsr-53-vietnamese").to(device) 
mt5 = pipelines.pipeline("text2text-generation","ontocord/mt5-fix-asr-vietnamese", device=0 if device == "cuda" else -1)
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.to(device), attention_mask=inputs.attention_mask.to(device)).logits
predicted_ids = torch.argmax(logits, dim=-1)
print("Prediction:", [aHash['generated_text'] for aHash in mt5(processor.batch_decode(predicted_ids), max_length=100)])
print("Reference:", test_dataset["sentence"][:2])

Evaluation

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

import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor, pipelines
import re
test_dataset = load_dataset("common_voice", "vi", split="test")
wer = load_metric("wer")

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
processor = Wav2Vec2Processor.from_pretrained("ontocord/wav2vec2-large-xlsr-vietnamese")
model = Wav2Vec2ForCTC.from_pretrained("ontocord/wav2vec2-large-xlsr-vietnamese").to(device)
mt5 = pipelines.pipeline("text2text-generation","ontocord/mt5-fix-asr-vietnamese", device=0 if device == "cuda" else -1)
chars_to_ignore_regex = '[\\\+\@\ǀ\,\?\.\!\-\;\:\"\“\%\‘\”\�]'
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)

# you may also want to use the  decode_string from https://huggingface.co/Nhut/wav2vec2-large-xlsr-vietnamese
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(device), attention_mask=inputs.attention_mask.to(device)).logits
    pred_ids = torch.argmax(logits, dim=-1)
    max_length = int(pred_ids.size()[1])
    txt = [aHash['generated_text'].strip() for aHash in mt5(processor.batch_decode(pred_ids), max_length=max_length)]
    batch["pred_strings"] = txt
  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: 25.207182

Training

The Common Voice train, validation, and FPT datasets were used for training. The script used for training can be found here # TODO