Wav2Vec2-BERT - Alvin

This model is a fine-tuned version of facebook/w2v-bert-2.0. This has a CER of 10.27 on Common Voice 16 (yue) test set (without punctuations).

Training and evaluation data

For training, three datasets were used:

• Common Voice 16 zh-HK and yue Train Set
• CantoMap: Winterstein, Grégoire, Tang, Carmen and Lai, Regine (2020) "CantoMap: a Hong Kong Cantonese MapTask Corpus", in Proceedings of The 12th Language Resources and Evaluation Conference, Marseille: European Language Resources Association, p. 2899-2906.
• Cantonse-ASR: Yu, Tiezheng, Frieske, Rita, Xu, Peng, Cahyawijaya, Samuel, Yiu, Cheuk Tung, Lovenia, Holy, Dai, Wenliang, Barezi, Elham, Chen, Qifeng, Ma, Xiaojuan, Shi, Bertram, Fung, Pascale (2022) "Automatic Speech Recognition Datasets in Cantonese: A Survey and New Dataset", 2022. Link: https://arxiv.org/pdf/2201.02419.pdf

Code Example

from transformers import pipeline
bert_asr = pipeline(
    "automatic-speech-recognition", model="alvanlii/wav2vec2-BERT-cantonese", device="cuda"
)
text = pipe(file)["text"]

or

import torch
import soundfile as sf
from transformers import AutoModelForCTC, Wav2Vec2BertProcessor

model_name = "alvanlii/wav2vec2-BERT-cantonese"

asr_model = AutoModelForCTC.from_pretrained(model_name).to(device)
processor = Wav2Vec2BertProcessor.from_pretrained(model_name)

audio_input, _ = sf.read(file)

inputs = processor([audio_input], sampling_rate=16_000).input_features
features = torch.tensor(inputs)

with torch.no_grad():
  logits = asr_model(features).logits

predicted_ids = torch.argmax(logits, dim=-1)
predictions = processor.batch_decode(predicted_ids, skip_special_tokens=True)

Training Hyperparameters

• learning_rate: 5e-5
• train_batch_size: 4 (on 1 3090)
• eval_batch_size: 1
• gradient_accumulation_steps: 32
• total_train_batch_size: 32x4=128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_warmup_steps: 1500