Nara SAR
Nara SAR is Nara's Burmese/English speech-recognition model, fine-tuned from Meta omniASR-CTC-300M. It supports Burmese, English and mixed speech, with offline CPU inference.
By Nara ยท naraaudio.app
This repository releases inference weights and inference source under Apache-2.0, with applicable third-party notices preserved. It does not contain training datasets or a complete training/reproduction pipeline.
Credit and model identity
When redistributing the model or a derivative, comply with Apache-2.0 sections 4(a)โ4(d): provide the license, retain applicable notices and attribution, mark modified files, and reproduce the applicable NOTICE attribution in an allowed location. Renaming an engine does not remove those obligations.
Suggested human-readable credit:
Powered by Nara SAR.
This short credit is suitable beside your app's engine label. Keep the full Nara/Meta attribution in the distributed legal notices or accompanying documentation, in a location permitted by the license. You do not need to repeat the full upstream chain beside every visible credit.
You may give your app or derivative its own name under Apache-2.0. Do not imply that your product is an official Nara release or endorsed by Nara. We ask that you identify the model's origin clearly rather than presenting an unchanged copy or format conversion as an independently trained model. See ATTRIBUTION.md and BRANDING.md.
The suggested credit sentence and UI label are recommendations, not additional Apache-2.0 license conditions. The NOTICE file preserves the actual attribution.
Inference files
| File | Purpose |
|---|---|
model.safetensors |
Full-precision checkpoint tensors, losslessly repackaged |
tokenizer.model |
Matching character tokenizer |
sar.json |
Character mapping and fixed final decoder settings |
language.lm |
Portable bilingual character 6-gram language model |
onnx/acoustic.onnx |
CPU INT8 MatMul export; convolutions remain FP32 |
onnx/acoustic-fp32.onnx |
Full-precision ONNX export |
sar_acoustic.py |
Standalone Torch inference architecture |
sar_decoder.py, sar_syllables.py |
Final beam decoder and structural rules |
transcribe.py |
Standalone local inference example |
SHA256SUMS, release_manifest.json |
Release integrity and format information |
Download only the backend you need plus the shared config, LM, decoder and runner files. A complete repository checkout includes multiple representations of the same acoustic model and is therefore larger than an INT8-only download. The INT8 graph is about 366 MB, the shared language model about 84 MB, and each full-precision representation about 1.3 GB (decimal sizes).
Quick start: offline CPU
Python 3.11 is the tested version. From a downloaded copy of this repository:
python -m pip install -r requirements.txt
python transcribe.py --audio speech.wav
The example loads local files and does not download weights or contact an ASR service. Use a PCM/float WAV or a format your installed SoundFile/libsndfile can read, such as FLAC. The supplied example accepts up to 30 minutes / 256 MiB, averages channels, resamples to 16 kHz and processes bounded chunks.
Full-precision ONNX:
python transcribe.py --audio speech.wav --precision fp32
Full-precision Torch checkpoint:
python -m pip install torch==2.8.0 --index-url https://download.pytorch.org/whl/cpu
python -m pip install -r requirements-torch.txt
python transcribe.py --audio speech.wav --backend torch
Use --json for structured output and --threads 8 to test a different CPU
thread count. The default is two threads. Tune it on your hardware; the fastest
setting and numerical output can differ across devices. There is no automatic
Transformers pipeline() integration in this release. Use the supplied runner.
Preprocessing and decoding matter
The checkpoint is the bilingual run3, step 6000 model. Input is mono 16 kHz,
normalized to zero mean and unit variance with epsilon 1e-5 per chunk.
Final output uses CTC prefix beam search with beam 16, top-k 10, alpha 0.6, beta 0, blank ID 0, the included bilingual character LM, and the included Burmese structural syllable constraints. Greedy argmax or omitting the LM/constraints is a different decoding recipe and can produce different text. The supplied decoder reports a failure rather than silently substituting greedy output. It does not return timestamps or diarization.
The runner uses sequential chunks of about 8โ10 seconds, preferring quiet split points. This is completed-audio transcription, not a streaming ASR model.
Validation and limitations
No independently held-out evaluation is reported in this release. We make no CER/WER, state-of-the-art, or superiority claim against Gemini or other ASR systems. Functional conversion checks and timing smoke tests are not accuracy benchmarks. See VALIDATION.md.
- Speech errors, omissions and hallucinated words remain possible. Review transcripts, especially names, numbers, accented English and mixed speech.
- Recordings, accents, noise conditions and devices beyond the small local test set are not comprehensively evaluated.
- INT8 quantization, CPU kernels, thread counts and chunk boundaries can change predictions. One local mixed-speech test changed an English word when the Android thread count changed; broader accuracy effects are unknown.
- This repository does not promise legal Burmese spelling for every output: the structural gate is only part of beam decoding, not a correctness proof.
- Do not use unreviewed transcripts as the sole basis for consequential decisions.
License and upstream work
The Nara model release uses Apache-2.0. The base model is Meta's omniASR-CTC-300M, also published under Apache-2.0. The portable Torch architecture follows fairseq2's MIT-licensed wav2vec2 implementation; its notice and license are included. Dependencies retain their own licenses. See THIRD_PARTY.md and NOTICE.
Apache-2.0 does not grant general rights to Nara or Meta names, logos or trademarks. Descriptive attribution is distinct from suggesting endorsement.
Citation
@misc{nara_sar_2026,
title = {Nara SAR: Burmese and English Speech Recognition},
author = {{Nara contributors}},
year = {2026},
url = {https://naraaudio.app}
}
Please also acknowledge Meta's Omnilingual ASR work.
- Downloads last month
- 32
Model tree for YeeNyunn/nara-sar
Base model
facebook/omniASR-CTC-300M