We did it! AI that writes Thai speech down as you talk, built in Thailand, free: OpenThai 2.0 ASR Streaming

We did it! AI that listens to Thai and writes it down while the speaker is still talking, built by a Thai team and free to download: OpenThai 2.0 ASR Streaming. Commercial use is allowed.
Sound familiar? You are watching a live stream, sitting in a class or in an online meeting, and you want live Thai captions. But most open Thai speech-to-text models have to hear a whole chunk first, and by the time the text appears the speaker has moved on.
Yesterday we released Thai document AI. Today is the second free release of the week.
- Text while you speak: just 0.08 to 1.12 seconds behind the voice. One model has four latency settings: pick speed or accuracy.
- 0.64B: one NVIDIA H100 keeps up with about 20 live streams (at 560 ms). No GPU? A CPU with 8 threads keeps up live (at 320 or 560 ms).
- Runs on your own machines, so audio never leaves the organisation. It suits organisations that cannot send audio to a foreign cloud, such as hospitals and banks.
- Trained on 12,373 hours of Thai speech, starting from NVIDIA Nemotron 3.5 ASR Streaming, which got 98.96% of Thai characters wrong before we fine-tuned it. It now gets 8.47% wrong on the same test set (FLEURS).
- The same model is behind iApp's OpenThai ASR Realtime service.
A direct comparison with Typhoon ASR Streaming 0.6B, built on the same NVIDIA base. We decoded both ourselves, on the same test sets, with the same script (numbers are the % of characters wrong, lower is better, at 560 ms):
- Read speech (FLEURS): 8.47 vs 11.28
- YouTube clips (GigaSpeech2): 15.35 vs 16.33
- Meetings (LOTUSDIS): 20.77 vs 28.91. This set favours us, because we trained on other meetings recorded in the same rooms.
Honestly: on TV talk shows Typhoon still does better (12.41 vs 10.84), and on the Isan dialect the two tie (34.90 vs 34.15). The default mode writes English words in Thai script, and th-mix writes only some of them in Latin letters. Long recordings must be cut at pauses before decoding, which the included server does for you. The team is working on it. Thanks to the Typhoon team for raising the bar for Thai speech recognition with us.
Why free? We want people who are deaf or hard of hearing, and older people, to be able to follow live streams, the news, classes and meetings with live Thai captions, anywhere. And anyone building a Thai voice assistant that responds quickly has to start by keeping up with Thai speech.
Download: huggingface.co/iapp/OpenThai2.0-ASR-Streaming, and the ONNX export (with int8): huggingface.co/iapp/OpenThai2.0-ASR-Streaming-ONNX Rather not run a GPU? Use the API: OpenThai ASR Realtime (free during the launch) Full results, usage and limits: OpenThai 2.0 ASR Streaming model page More on the iApp site: the launch post and the model page
Built on NVIDIA Nemotron · trained on NVIDIA H100 GPUs supported by Siam AI
