transformers/README.md

Transformers FP32 export

This export uses the standard ParakeetForTDT implementation in Transformers 5.15.1. No custom model code or trust_remote_code is required. Root-level configuration, processor and safetensors files let applications such as Buzz load oruk/orukeet through their existing Hugging Face/Parakeet backend.

pip install 'torch==2.8.0' 'torchaudio==2.8.0' 'transformers==5.15.1' soundfile

import soundfile as sf
from transformers import AutoModelForTDT, AutoProcessor, pipeline

model = AutoModelForTDT.from_pretrained("oruk/orukeet", dtype="float32")
processor = AutoProcessor.from_pretrained("oruk/orukeet")
asr = pipeline(
    "automatic-speech-recognition", model=model, device="cpu",
    tokenizer=processor.tokenizer, feature_extractor=processor.feature_extractor,
)
audio, sample_rate = sf.read("speech.wav", dtype="float32")
assert sample_rate == 16000 and audio.ndim == 1
print(asr(audio)["text"])

Use 16 kHz mono audio. The model detects language automatically and does not provide translation, speaker labels or prompt-conditioned transcription. Buzz currently returns 30-second chunk boundaries, not aligned word timestamps. Downloading the required configuration uses Hugging Face's normal model download accounting. Cached local loading does not require an extra tracking request.

Conversion

The source is orukeet-v0.1.0.nemo at revision 555136b50265a132d4cea0d35560c26fc4f657ab, SHA-256 031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56. All 723 mapped tensors load with strict=True: 627,008,134 parameters in FP32. The materialized Gabor filters are preserved as ordinary convolution weights.

The bundled upstream converter comes from Transformers commit 6c6bac29f50c8aad5d1f06c72b2f892a6e335dd0. The wrapper verifies the source hash, extracts it with the data safety filter, and requires exact state-dictionary and tokenizer compatibility before saving. Reproduce in Python 3.13:

pip install 'torch==2.8.0' 'transformers==5.15.1' protobuf sentencepiece pyyaml
python convert.py /path/to/output

Download both convert.py and convert_nemo_to_hf.py from this directory. manifest.json records every exported runtime file's hash and size. Weights remain CC BY-SA 4.0 with NVIDIA attribution; see the root NOTICE.md and LICENSE-WEIGHTS. The upstream converter is Apache-2.0; see LICENSE-CONVERTER.

Qualification

On Apple M5 Max CPU with PyTorch 2.8.0 and Transformers 5.15.1:

• The actual Buzz transcription path passed English, German, Spanish and French repeated transcription, 0.1/1/5-second silence, a recording spanning two chunks, and local loading/inference with socket connections disabled.
• A fixed 120-clip, six-language FLEURS validation compatibility sample produced the same normalized transcript as the recorded source-NeMo output for 117 clips. Pooled WER was 5.43% for this export versus 5.34% for source NeMo using the same normalization: two additional word errors across 2,433 reference words.
• The sample contained 1,257.82 seconds of audio and took 217.64 seconds of inference across eight CPU threads (about 0.17 real-time factor). Other work ran on the machine during this check; this is a runtime receipt, not a controlled speed comparison with Parakeet.

This small compatibility sample is not a new held-out benchmark. The published 25-language NeMo results in the model card do not automatically transfer to this export. Per-language measurements and normalization are in qualification.json.