The best Vietnamese speech recognition using Conforme-CTC 2024

https://github.com/thinhlx1993/vietnamese_asr

Training data

I collect data from many different sources for training. The training data contains over 10k hours of speech data from the sources below

• Common Voice dataset
• VIVOS dataset
• (AN4) database audio files
• Vietnamese Speech recognition
• Youtube public dataset
• Vietnamese Dialogue Telephony speech dataset
• Travel Call Center Speech Data
• LibriSpeech

Models setup

Tokenizer SentencePieceTokenizer initialized with 128 tokens

121 M Total params

Name	Type	Params
0	preprocessor	AudioToMelSpectrogramPreprocessor
1	encoder	ConformerEncoder
2	decoder	ConvASRDecoder
3	loss	CTCLoss
4	spec_augmentation	SpectrogramAugmentation
5	wer	WER

Benchmark WER result

	WER	CER
without ngram LM	10.71	12.21
with ngram LM	9.15	10.2

How to use

Download model weight here

https://drive.google.com/drive/folders/1SVNibfeMshfVkmatIU90LYok_Mf0zMD0?usp=sharing

Install Nemo Frameworks

https://github.com/NVIDIA/NeMo

You can try demo in the example folder or this one

I created a free-to-use API server to submit the inference data

The file input should have a bitrate of 16000 to avoid hidden bugs

File duration must be lower than 10s

import subprocess

command = [
    "curl", "--location", "https://api.voicesplitter.com/api/v1/uploads",
    "--form", 'file=@"/path/to/your/wav_file.wav"'
]

result = subprocess.run(command, capture_output=True, text=True)
print(result.stdout.encode('utf-8').decode('unicode_escape'))

Contact

[email protected] | Thinh Le's LinkedIn Profile