NbAiLab/vibevoice-asr-north-saami-v5

North Saami ASR model obtained by fine-tuning VibeVoice ASR on nb-asr-parakeet v5 audio/text pairs.

Run Info

Run name: olivia_vibevoice_asr_v5_full_lr4e-6_e100_h200x4-full-evalstart-wu1000-eval256-flash-attention-3-ds_zero3_bf16_550392
Base model: microsoft/VibeVoice-ASR
Dataset: nb-asr-parakeet data_v5
Local artifact: /cluster/work/projects/nn30001k/versae/nb-vibevoice/outputs/olivia_vibevoice_asr_v5_full_lr4e-6_e100_h200x4-full-evalstart-wu1000-eval256-flash-attention-3-ds_zero3_bf16_550392
Weights & Biases: https://wandb.ai/nbailab/nb-sami-asr-north-vibevoice/runs/w550392
Attention backend: flash_attention_3
Fine-tune mode: full
Freeze speech tokenizers: True
Num epochs: 100.0
Per-device train batch size: 1
Gradient accumulation steps: 4
Learning rate: 4e-06
DeepSpeed config: /cluster/projects/nn30001k/versae/VibeVoice/finetuning-asr/ds_zero3_bf16.json

The uploaded folder contains the final saved model and processor artifacts.
Intermediate checkpoint-* folders are intentionally omitted from the model repo upload.
Tokenizer audit output is included when available.

Safetensors

Model size

9B params

Tensor type

BF16