giangndm/parakeet-ctc-0.6b-vietnamese-bf16

Parakeet CTC 0.6B Vietnamese (BF16)

Transformers-compatible BF16 conversion of NVIDIA's NeMo checkpoint nvidia/parakeet-ctc-0.6b-Vietnamese. It contains a ParakeetForCTC model, its processor/tokenizer, and one model.safetensors file. It does not use remote code.

The source model's license and use restrictions apply. This conversion does not make any claim of changed ASR quality.

CTC recognition

import torch
from transformers import AutoModelForCTC, AutoProcessor

repo_id = "giangndm/parakeet-ctc-0.6b-vietnamese-bf16"
processor = AutoProcessor.from_pretrained(repo_id)
model = AutoModelForCTC.from_pretrained(repo_id, dtype=torch.bfloat16).cuda().eval()

inputs = processor(waveform_16khz, sampling_rate=16_000, return_tensors="pt")
inputs = inputs.to("cuda", dtype=torch.bfloat16)
with torch.inference_mode():
    token_ids = model.generate(**inputs)
print(processor.decode(token_ids[0]))

Encoder features

import torch
from transformers import AutoProcessor, ParakeetEncoder

processor = AutoProcessor.from_pretrained(repo_id)
encoder = ParakeetEncoder.from_pretrained(repo_id, dtype=torch.bfloat16).cuda().eval()
inputs = processor(waveform_16khz, sampling_rate=16_000, return_tensors="pt")
inputs = inputs.to("cuda", dtype=torch.bfloat16)
with torch.inference_mode():
    output = encoder(**inputs)

features = output.last_hidden_state
frame_mask = output.attention_mask

The encoder consumes 80-bin log-mel features made by ParakeetFeatureExtractor and emits 1024-dimensional hidden states after 8x time subsampling.