Wav2Vec2-Large-XLSR-53-Cantonese

Fine-tuned facebook/wav2vec2-large-xlsr-53 on Cantonese using the Common Voice Corpus 8.0. When using this model, make sure that your speech input is sampled at 16kHz.

The Common Voice's validated train and dev were used for training.

The script used for training can be found at https://github.com/holylovenia/wav2vec2-pretraining.

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", "zh-HK", split="test[:2%]")

processor = Wav2Vec2Processor.from_pretrained("CAiRE/wav2vec2-large-xlsr-53-cantonese")
model = Wav2Vec2ForCTC.from_pretrained("CAiRE/wav2vec2-large-xlsr-53-cantonese")


# 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"])
  resampler = torchaudio.transforms.Resample(sampling_rate, 16_000)
  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 zh-HK 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", "zh-HK", split="test")
wer = load_metric("cer")

processor = Wav2Vec2Processor.from_pretrained("CAiRE/wav2vec2-large-xlsr-53-cantonese")
model = Wav2Vec2ForCTC.from_pretrained("CAiRE/wav2vec2-large-xlsr-53-cantonese") 
model.to("cuda")

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


# 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"])
  resampler = torchaudio.transforms.Resample(sampling_rate, 16_000)
  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("CER: {:2f}".format(100 * cer.compute(predictions=result["pred_strings"], references=result["sentence"])))

Test Result: CER: 18.55 %

Citation

If you use our code/model, please cite us:

@inproceedings{lovenia2022ascend,
  title={ASCEND: A Spontaneous Chinese-English Dataset for Code-switching in Multi-turn Conversation},
  author={Lovenia, Holy and Cahyawijaya, Samuel and Winata, Genta Indra and Xu, Peng and Yan, Xu and Liu, Zihan and Frieske, Rita and Yu, Tiezheng and Dai, Wenliang and Barezi, Elham J and others},
  booktitle={Proceedings of the 13th Language Resources and Evaluation Conference (LREC)},
  year={2022}
}