Downloads · 30 days
0
parakeet-tdt-lt-onnx
aai-labs/parakeet-tdt-lt-onnx
parakeet-tdt-lt-onnx is a automatic speech recognition model from aai-labs. Use it when you need speech turned into text. The card lists the license as cc-by-4.0.
ONNX (int8) export of sliderforthewin/parakeet-tdt-lt — a Lithuanian fine-tune of nvidia/parakeet-tdt-0.6b-v3 — for use with sherpa-onnx (modeltype="nemotransducer", featuredim=128, greedy search, CPU).
Access
Public
Updated Jul 24, 2026
Repo size
671 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.onnx670 MB · 100%
At a glance
- Task
- Automatic Speech Recognition
- License
- cc-by-4.0
- Access
- Public
- Created
- Jul 24, 2026
- Updated
- Jul 24, 2026
- SHA
- 9fedf81f
Base models
Model card
From the Hugging Face model README
parakeet-tdt-lt-onnx
ONNX (int8) export of sliderforthewin/parakeet-tdt-lt
— a Lithuanian fine-tune of nvidia/parakeet-tdt-0.6b-v3 — for use with
sherpa-onnx (model_type="nemo_transducer",
feature_dim=128, greedy search, CPU).
Files
| file | size | purpose |
|---|---|---|
encoder.int8.onnx | 622 MB | Conformer encoder (int8) |
decoder.int8.onnx | 11 MB | TDT prediction network (int8) |
joiner.int8.onnx | 6 MB | TDT joiner (int8) |
tokens.txt | 100 KB | SentencePiece vocab (8192 + <blk>) |
parakeet-lt-test.wav | 0.6 MB | FLEURS-LT dev clip (CC-BY 4.0) for smoke tests |
Usage
// sherpa-onnx C++ (cxx-api)
OfflineRecognizerConfig cfg;
cfg.feat_config.sample_rate = 16000;
cfg.feat_config.feature_dim = 128; // v3 preprocessor
cfg.model_config.transducer.encoder = "encoder.int8.onnx";
cfg.model_config.transducer.decoder = "decoder.int8.onnx";
cfg.model_config.transducer.joiner = "joiner.int8.onnx";
cfg.model_config.tokens = "tokens.txt";
cfg.model_config.model_type = "nemo_transducer";
cfg.model_config.provider = "cpu";
cfg.decoding_method = "greedy_search";
Reference implementation: aai-labs meeting transcriber
(src/parakeet_transcriber.cpp). Export recipe: tools/export-parakeet/ there
(adapted from k2-fsa's official scripts/nemo/parakeet-tdt-0.6b-v3).
Known quirk: greedy decoding can flip between diacritic and plain-ASCII styling depending on runtime int8 numerics (model property of the fine-tune; beam + the author's n-gram LM stabilizes both WER and styling).
License: CC-BY-4.0 (same as the source checkpoint and training data).
- Task
- Automatic Speech Recognition
- License
- cc-by-4.0
- Created
- Jul 24, 2026
- Updated
- Jul 24, 2026