nmcuong/parakeet-tdt-0.6b-v3-vietnamese

Parakeet-TDT-0.6B Vietnamese

Fine-tuned NVIDIA Parakeet-TDT-0.6B-v3 for Vietnamese and English automatic speech recognition.

Model Description

• Architecture: FastConformer encoder + Token-and-Duration Transducer (TDT) decoder
• Parameters: 0.6B
• Tokenizer: 8,192-token SentencePiece BPE
• Base model: nvidia/parakeet-tdt-0.6b-v3
• Languages: Vietnamese (primary), English
• Training data: 8,750h Vietnamese + 1,530h English (~10,280h total)

Evaluation Results

Test Set	WER
Validation set	5.68%

Usage

import nemo.collections.asr as nemo_asr

# Load model
model = nemo_asr.models.ASRModel.from_pretrained("nmcuong/parakeet-tdt-0.6b-v3-vietnamese")

# Transcribe
transcriptions = model.transcribe(["audio.wav"])
print(transcriptions[0].text)

Training Details

• Base model: nvidia/parakeet-tdt-0.6b-v3
• Learning rate: 5e-5
• Batch size: 48 (gradient accumulation: 2, effective batch size: 96)
• Training steps: 130,000
• GPUs: 2× H200
• Precision: bf16-mixed
• Audio duration filter: min 0.1s, max 30s
• Max WER filter: 12.5%

Training Data

Dataset	Duration
Vietnamese	8,750h
English	1,530h
Total	~10,280h

Acknowledgments

• Base model: NVIDIA Parakeet-TDT-0.6B-v3