language	datasets	tags	license
th	common_voice	audio automatic-speech-recognition speech xlsr-fine-tuning	cc-by-sa 4.0

wav2vec2-large-xlsr-53-th

Finetuning wav2vec2-large-xlsr-53 on Thai Common Voice 7.0

Read more on our blog

We finetune wav2vec2-large-xlsr-53 based on Fine-tuning Wav2Vec2 for English ASR using Thai examples of Common Voice Corpus 7.0. The notebooks and scripts can be found in vistec-ai/wav2vec2-large-xlsr-53-th. The pretrained model and processor can be found at airesearch/wav2vec2-large-xlsr-53-th.

Usage

#load pretrained processor and model
processor = Wav2Vec2Processor.from_pretrained("airesearch/wav2vec2-large-xlsr-53-th")
model = Wav2Vec2ForCTC.from_pretrained("airesearch/wav2vec2-large-xlsr-53-th")

#function to resample to 16_000
def speech_file_to_array_fn(batch, 
                            text_col="sentence", 
                            fname_col="path",
                            resampling_to=16000):
    speech_array, sampling_rate = torchaudio.load(batch[fname_col])
    resampler=torchaudio.transforms.Resample(sampling_rate, resampling_to)
    batch["speech"] = resampler(speech_array)[0].numpy()
    batch["sampling_rate"] = resampling_to
    batch["target_text"] = batch[text_col]
    return batch

#get 2 examples as sample input
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)

#infer
with torch.no_grad():
    logits = model(inputs.input_values,).logits

predicted_ids = torch.argmax(logits, dim=-1)

print("Prediction:", processor.batch_decode(predicted_ids))
print("Reference:", test_dataset["sentence"][:2])

>> Prediction: ['และ เขา ก็ สัมผัส ดีบุก', 'คุณ สามารถ รับทราบ เมื่อ ข้อความ นี้ ถูก อ่าน แล้ว']
>> Reference: ['และเขาก็สัมผัสดีบุก', 'คุณสามารถรับทราบเมื่อข้อความนี้ถูกอ่านแล้ว']

Datasets

Common Voice Corpus 7.0](https://commonvoice.mozilla.org/en/datasets) contains 133 validated hours of Thai (255 total hours) at 5GB. We pre-tokenize with pythainlp.tokenize.word_tokenize. We preprocess the dataset using cleaning rules described in notebooks/cv-preprocess.ipynb by @tann9949. We then deduplicate and split as described in ekapolc/Thai_commonvoice_split in order to 1) avoid data leakage due to random splits after cleaning in Common Voice Corpus 7.0 and 2) preserve the majority of the data for the training set. The dataset loading script is scripts/th_common_voice_70.py. You can use this scripts together with train_cleand.tsv, validation_cleaned.tsv and test_cleaned.tsv to have the same splits as we do. The resulting dataset is as follows:

DatasetDict({
    train: Dataset({
        features: ['path', 'sentence'],
        num_rows: 86586
    })
    test: Dataset({
        features: ['path', 'sentence'],
        num_rows: 2502
    })
    validation: Dataset({
        features: ['path', 'sentence'],
        num_rows: 3027
    })
})

Training

We fintuned using the following configuration on a single V100 GPU and chose the checkpoint with the lowest validation loss. The finetuning script is scripts/wav2vec2_finetune.py

# create model
model = Wav2Vec2ForCTC.from_pretrained(
    "facebook/wav2vec2-large-xlsr-53",
    attention_dropout=0.1,
    hidden_dropout=0.1,
    feat_proj_dropout=0.0,
    mask_time_prob=0.05,
    layerdrop=0.1,
    gradient_checkpointing=True,
    ctc_loss_reduction="mean",
    pad_token_id=processor.tokenizer.pad_token_id,
    vocab_size=len(processor.tokenizer)
)
model.freeze_feature_extractor()
training_args = TrainingArguments(
    output_dir="../data/wav2vec2-large-xlsr-53-thai",
    group_by_length=True,
    per_device_train_batch_size=32,
    gradient_accumulation_steps=1,
    per_device_eval_batch_size=16,
    metric_for_best_model='wer',
    evaluation_strategy="steps",
    eval_steps=1000,
    logging_strategy="steps",
    logging_steps=1000,
    save_strategy="steps",
    save_steps=1000,
    num_train_epochs=100,
    fp16=True,
    learning_rate=1e-4,
    warmup_steps=1000,
    save_total_limit=3,
    report_to="tensorboard"
)

Evaluation

We benchmark on the test set using WER with words tokenized by PyThaiNLP 2.3.1 and deepcut, and CER. We also measure performance when spell correction using TNC ngrams is applied. Evaluation codes can be found in notebooks/wav2vec2_finetuning_tutorial.ipynb. Benchmark is performed on test-unique split.

	WER PyThaiNLP 2.3.1	WER deepcut	CER
Kaldi from scratch	23.04		7.57
Ours without spell correction	13.634024	8.152052	2.813019
Ours with spell correction	17.996397	14.167975	5.225761
Google Web Speech API※	13.711234	10.860058	7.357340
Microsoft Bing Speech API※	12.578819	9.620991	5.016620
Amazon Transcribe※	21.86334	14.487553	7.077562
NECTEC AI for Thai Partii API※	20.105887	15.515631	9.551027

※ APIs are not finetuned with Common Voice 7.0 data

License

Models are under CC-BY-SA 4.0. All codes are under Apache 2.0.

Ackowledgements

• model training and validation notebooks/scripts @cstorm125
• dataset cleaning scripts @tann9949
• dataset splits @ekapolc and @14mss
• running the training @mrpeerat
• spell correction @wannaphong