Downloads · 30 days
0
Nawanop/typhoon-asr-realtime-safetensors
typhoon-asr-realtime-safetensors is a automatic speech recognition model from Nawanop. Use it when you need speech turned into text. It is set up for nemo. The card lists the license as cc-by-4.0.
safetensors re-export of typhoon-ai/typhoon-asr-realtime — a Thai real-time ASR model (NVIDIA FastConformer-Transducer, NeMo) by SCB10X / OpenTyphoon.
Downloads · 30 days
0
Access
Public
Updated Jul 16, 2026
Parameters
115M
693 MB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors693 MB · 100%
From the Hugging Face model README
safetensors re-export of
typhoon-ai/typhoon-asr-realtime —
a Thai real-time ASR model (NVIDIA FastConformer-Transducer, NeMo) by SCB10X / OpenTyphoon.
This repository contains only a format conversion of the original weights. No retraining, no finetuning, no architecture change. Two precision variants are provided.
⚠️ This is an unofficial community re-export, not published by the original authors. For the canonical model, always refer to typhoon-ai/typhoon-asr-realtime.
| File | Precision | Size | Notes |
|---|---|---|---|
model.safetensors | fp32 | ~440 MB | Full precision. Bit-exact with the original checkpoint. Use this for CPU inference or further training. |
model.fp16.safetensors | fp16 | ~220 MB | Half precision, weights cast from fp32. Half the size. Best for GPU inference / distribution. On CPU it is cast back to fp32 at load time (no speedup, slight accuracy loss) — prefer the fp32 file for CPU. |
config.yaml | — | — | NeMo architecture + preprocessing config. |
tokenizer/ | — | — | SentencePiece tokenizer (model + vocab). |
Parameters: ~115.4M. Sample rate: 16 kHz mono.
Load the weights into the base NeMo architecture:
import nemo.collections.asr as nemo_asr
from safetensors.torch import load_model
# instantiate the original architecture, then overwrite weights from safetensors
model = nemo_asr.models.ASRModel.from_pretrained("typhoon-ai/typhoon-asr-realtime")
load_model(model, "model.safetensors") # or "model.fp16.safetensors"
model.eval()
print(model.transcribe(["audio.wav"])[0])
For the fp16 file, match the dtype before loading; on CPU, cast back to fp32:
model = nemo_asr.models.ASRModel.from_pretrained("typhoon-ai/typhoon-asr-realtime")
model = model.half()
load_model(model, "model.fp16.safetensors")
model = model.float() # skip this line if running on GPU
model.eval()
typhoon-ai/typhoon-asr-realtime, revision 2c58a30ba9a3bf92d095a5df91bec6996f04c3a1.nemo checkpoint with NeMo, serialized state_dict() via safetensors.torch.save_model. fp16 file produced with model.half() before serialization.nemo_toolkit 2.7.3, torch 2.13.0, safetensors 0.8.0model.safetensors: d915c3e43c1486b76068bccf10752fec9d30bc34279cf25b7de1909591661781model.fp16.safetensors: 17b1ea54bb381a6e52f6d7596664a8470b40522b7a0038054b4b53afc84abcb6The fp32 export is bit-exact with the original weights (verified: 705/705 tensors equal).
This is a derivative work of typhoon-ai/typhoon-asr-realtime, released by SCB10X / OpenTyphoon under the Creative Commons Attribution 4.0 International (CC-BY-4.0) license. This re-export is distributed under the same CC-BY-4.0 license.
Per CC-BY-4.0, this derivative:
safetensors and an fp16 precision
cast. No changes to the trained weights (fp32) or architecture.Use of this model is also subject to the original OpenTyphoon Terms & Conditions and Privacy Notice:
@misc{warit2026typhoonasr,
title={Typhoon ASR Real-time: FastConformer-Transducer for Thai Automatic Speech Recognition},
author={Warit Sirichotedumrong and Adisai Na-Thalang and Potsawee Manakul and Pittawat Taveekitworachai and Sittipong Sripaisarnmongkol and Kunat Pipatanakul},
year={2026},
eprint={2601.13044},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2601.13044}
}
Provided "as is", without warranty of any kind. This re-export is not endorsed by or affiliated with SCB10X / OpenTyphoon. All rights to the underlying model remain with the original authors. If the original authors request removal, this repository will be taken down.