nectec/Pathumma-stt-th-fastconformer-rnnt-large

🪷 Pathumma STT Fastconformer RNNT Large (TH)

Pathumma STT Fastconformer RNNT Large (TH) is a Thai-English ASR model fine-tuned for robust, real-world speech recognition, including background noise, real-time inference, and Thai-English code-switching.

The model is trained on a combination of open-source and in-house datasets, covering general, news, technology, science, finance, medical & healthcare, and legal domains.

The model is built on NVIDIA's FastConformer-RNNT architecture, which supports low-latency, real-time inference.

NVIDIA NeMo

To train, fine-tune, or run Python inference with this model, install NVIDIA NeMo after installing a recent PyTorch version.

pip install -U nemo_toolkit['asr']

Quickstart

Load the model:

import torch
import nemo.collections.asr as nemo_asr

device = "cuda" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.bfloat16 if torch.cuda.is_available() else torch.float32

model_name = "nectec/Pathumma-stt-th-fastconformer-rnnt-large"
asr_model = nemo_asr.models.EncDecRNNTBPEModel.from_pretrained(model_name=model_name)
asr_model.to(device)
asr_model.to(torch_dtype)

Run inference:

output = asr_model.transcribe(["audio_path.wav"])
print(output[0].text)

Run inference with timestamps:

output = asr_model.transcribe(["audio_path.wav"], timestamps=True)

# Output indices correspond to the order of items in `data`
word_timestamps = output[0].timestamp['word']
segment_timestamps = output[0].timestamp['segment']
char_timestamps = output[0].timestamp['char']

for word in word_timestamps:
    print(f"{word['start']}s - {word['end']}s : {word['segment']}")

Contributors

Wayupuk Sommuang, Siwakorn Kaewwichai, Sakson Bunta, Pattara Tipaksorn

Acknowledgements

We gratefully acknowledge the Pathumma LLM Team for their contributions and support, and ThaiSC (NSTDA Supercomputer Centre) for providing access to the LANTA supercomputer, which was used for model training, fine-tuning, and evaluation.

Citation

@misc{Sommuang2026PathummaFastconRNNT,
    title        = { {Pathumma STT Fastconformer RNNT Large (TH)} },
    author       = { 
      Wayupuk Sommuang and 
      Siwakorn Kaewwichai and
      Sakson Bunta and
      Pattara Tipaksorn
    },
    url          = { https://huggingface.co/nectec/Pathumma-stt-th-fastconformer-rnnt-large },
    publisher    = { Hugging Face },
    year         = { 2026 },
}