avencera/scriptrs-models

scriptrs models

Model files for scriptrs, a Rust library for fast NVIDIA Parakeet TDT v2 and v3 speech-to-text on Core ML (macOS) and native CUDA (Linux), at the same word error rate as NVIDIA's reference implementation. With the default online feature, scriptrs downloads these files on first use from a pinned revision of this repository.

These are format conversions of NVIDIA's models, not new or retrained models. The native CUDA weights are NVIDIA's FP32 weights with their original names; scriptrs converts selected layers to FP16 on the GPU at load time.

Same accuracy, faster

scriptrs runs NVIDIA's weights unchanged and gets the same word error rate as NVIDIA's own NeMo implementation, at 3x its speed. All results are Parakeet v3 on an RTX 5090.

LibriSpeech test-clean, English (5.4 h), one utterance at a time:

	Speed (x real time)	vs NeMo	WER
scriptrs CUDA	1534x	3.2x	1.93 %
NVIDIA NeMo (PyTorch)	477x	1x	1.93 %
parakeet-rs 0.4.3 (ONNX Runtime)	253x	0.53x	1.93 %

Multilingual

The multilingual check is nearly 6 times the size of LibriSpeech: 31.6 h over ten language sets in six languages, one utterance at a time with automatic language selection. In scriptrs's parity tests, its FP32 mode matches NeMo layer by layer and gives NeMo's exact text. The FP16 default scores the same as FP32:

Set	Speed, FP16 default	WER, FP16 default	WER, FP32
FLEURS de, es, fr, it, nl, pt (15.8 h)	1839x	4.57 %	4.57 %
MLS es, fr, it, pt (15.8 h)	1921x	6.25 %	6.26 %

Batches and long recordings

Workload	Speed (x real time)	WER
LibriSpeech test-clean, batches of 16	4767x	1.93 %, unchanged
Earnings-21, 44 calls (39 h), long-form	3633x	8.24 %

On Earnings-21, NeMo transcribing each whole call with local attention scores 9.58 %.

Against Whisper large-v3

On harder English audio, long-form with the defaults and English selected has a lower WER than Whisper large-v3 (run through WhisperX 3.8.6 in float16) on every set, and runs 10x to 14x as fast:

Set	Recordings	Hours	scriptrs WER	Whisper WER	scriptrs speed	Whisper speed	vs Whisper
Earnings-21 earnings calls	44	39.3	8.2 %	9.7 %	2458x	186x	13.2x
AMI meetings, headset mix	34	18.7	21.5 %	23.8 %	2420x	182x	13.3x
AMI meetings, far-field microphone	34	18.7	23.2 %	28.3 %	2414x	199x	12.1x
EdAcc accented conversations	40	18.2	16.0 %	19.2 %	2433x	201x	12.1x
CHiME-6 dinner parties	4	9.7	44.0 %	44.7 %	2438x	176x	13.8x
AppTek call-center calls, 14 accents	56	9.4	11.2 %	12.6 %	2220x	224x	9.9x
Mean over sets	212	114	20.7 %	23.1 %	2417x	191x	12.6x

WER is pooled over each set's recordings, both sides through Whisper's English normalizer. Speed is x real time for transcription alone on one RTX 5090, with models loaded and audio in memory; WhisperX runs its VAD with batch size 16. The last row's speed is over all 114 hours.

More results are in the scriptrs README.

Usage

# macOS, Core ML
scriptrs = "0.3"

# Linux, NVIDIA GPU
scriptrs = { version = "0.3", features = ["cuda"] }

use scriptrs::{ModelBundle, ParakeetDeployment, TranscriptionPipeline};

// Core ML on macOS
let pipeline = TranscriptionPipeline::from_pretrained()?;

// native CUDA on Linux
let bundle = ModelBundle::from_pretrained_native(ParakeetDeployment::v3())?;
let pipeline = TranscriptionPipeline::from_bundle(bundle)?;

let result = pipeline.run(&audio)?; // mono 16 kHz f32 samples

For offline setups, download this repository and set SCRIPTRS_MODELS_DIR to its local path.

Contents

Path	Runtime	Source
parakeet-v2/encoder.mlmodelc, decoder.mlmodelc, joint-decision.mlmodelc	Core ML	nvidia/parakeet-tdt-0.6b-v2
parakeet-v3/encoder.mlmodelc, decoder.mlmodelc, joint-decision.mlmodelc	Core ML	nvidia/parakeet-tdt-0.6b-v3, via FluidInference/parakeet-tdt-0.6b-v3-coreml
parakeet-v2/model.safetensors	CUDA	nvidia/parakeet-tdt-0.6b-v2, exported from the NeMo checkpoint
parakeet-v3/model.safetensors	CUDA	nvidia/parakeet-tdt-0.6b-v3, exported from the NeMo checkpoint
parakeet-v2/vocab.txt, parakeet-v3/parakeet_vocab.json	both	NVIDIA tokenizers
vad/silero-vad.mlmodelc	Core ML	Silero VAD 5.1.2, via aufklarer/Silero-VAD-v5-CoreML
vad/silero-vad-v6.mlmodelc	Core ML	Silero VAD 6.2.3, via aufklarer/Silero-VAD-v6.2.1-CoreML, which has the same weights
vad/silero-vad-v5.safetensors	CUDA	Silero VAD 5.1.2, 16 kHz branch

The native CUDA files also live in model-named repositories, which scriptrs 0.3.1 and later download from: avencera/parakeet-tdt-0.6b-v2-cuda, avencera/parakeet-tdt-0.6b-v3-cuda and avencera/silero-vad-v5-safetensors, and from scriptrs 0.5.0 also avencera/silero-vad-v6-safetensors. The copies here serve scriptrs 0.3.0.

The native weight files record their source and its SHA-256 in the safetensors metadata. PROVENANCE.json and PROVENANCE-native.json list sources, revisions and checksums.

License

The Parakeet files are derived from NVIDIA Parakeet TDT 0.6B v2 and v3, licensed under CC BY 4.0. Credit: NVIDIA. Changes: converted to Core ML and to a safetensors layout for scriptrs. The Silero VAD files are derived from Silero VAD, licensed under the MIT License. See LICENSE.