nb-sami-asr-north-parakeet-e20-lr1p25e-3-g4

nb-asr-north-sami-parakeet-e20-lr1p25e-3-g4

North Sami ASR model (NeMo .nemo).

• Source run: olivia_parakeet_v6_full_v5_lr1p25e-3_e20_gh200x4-apptainer_441952
• Best val WER checkpoint observed: 0.1972
• Epochs: 20
• Learning rate: 1.25e-3
• GPUs: 4x GH200

Model was repacked so model_config.yaml contains name: nb-asr-north-sami-parakeet-e20-lr1p25e-3-g4.

Portable parakeet.cpp executable

nb-sami-asr-north-parakeet-e20-lr1p25e-3-g4-q5_k.parakeet is a self-contained CPU executable containing the Q5_K GGUF model and parakeet.cpp, packaged with Cosmopolitan Libc. It targets x86-64 and ARM64 and is intended to run across Linux, macOS, Windows, and BSD without a Python or NeMo installation.

On Linux and macOS:

chmod +x nb-sami-asr-north-parakeet-e20-lr1p25e-3-g4-q5_k.parakeet
./nb-sami-asr-north-parakeet-e20-lr1p25e-3-g4-q5_k.parakeet \
  --input audio.wav \
  --threads 8

If zsh on macOS reports exec format error, invoke it through the system shell:

sh -c './nb-sami-asr-north-parakeet-e20-lr1p25e-3-g4-q5_k.parakeet --input audio.wav --threads 8'

On Apple Silicon, the first launch compiles Cosmopolitan's embedded APE loader and therefore requires the Xcode Command Line Tools. Install them with:

xcode-select --install

Optional word timestamps:

./nb-sami-asr-north-parakeet-e20-lr1p25e-3-g4-q5_k.parakeet \
  --input audio.wav \
  --timestamps

The embedded model uses Q5_K quantization. Artifact SHA-256:

dc395d2c07d359c6ebe82fdbfb31407c9f6ba28c0fa465d09024140f346b948c

BF16 portable executable

nb-sami-asr-north-parakeet-e20-lr1p25e-3-g4-bf16.parakeet embeds a GGUF with 219 allowlisted matrix weights in BF16 and the remaining 482 tensors in F32. The F32 tensors include convolutions, LSTM weights, embeddings, normalization, biases, the output projection, and audio preprocessing buffers required by the engine's direct float access. This is a mixed BF16/F32 model.

The Cosmopolitan executable includes x86-64 and ARM64 engines and the Apple Silicon APE loader. Verified on Linux ARM64: exact transcript agreement with NeMo 2.7.3 on the first three clips (22.62 seconds) from shard 0 of the North Sámi test split, valid JSON timestamps, ZIP integrity, and direct and Bash launch. This is a smoke test, not a full-dataset accuracy evaluation. This BF16 artifact has not been executed on macOS, Windows, or x86-64 during this validation.

Download and run on Linux or macOS:

curl -L --fail --retry 3 \
  -o nb-sami-asr-north-parakeet-e20-lr1p25e-3-g4-bf16.parakeet \
  'https://huggingface.co/NbAiLab/nb-asr-north-sami-parakeet-e20-lr1p25e-3-g4/resolve/main/nb-sami-asr-north-parakeet-e20-lr1p25e-3-g4-bf16.parakeet?download=true'
chmod +x nb-sami-asr-north-parakeet-e20-lr1p25e-3-g4-bf16.parakeet
bash ./nb-sami-asr-north-parakeet-e20-lr1p25e-3-g4-bf16.parakeet \
  --input audio.wav --threads 8

On Apple Silicon, install the Xcode Command Line Tools (xcode-select --install) for the embedded loader's first launch. Add --timestamps for word timestamps, or --json for machine-readable output.

Artifact size: 1,447,884,065 bytes. SHA-256:

99f147ba2d24d7f094cfc71ab40d09418522b9a603c29ad00eddf5114aff3061