Skip to main content

Open weights

OpenThai 2.0 ASR Streaming

An open model that writes Thai speech down while people are still talking.

  • Parameters0.64billion
  • LicenceOpenMDW-1.1
  • Text behind the voice80–1120ms
  • FLEURS, characters wrong8.47%

Download the weightsRead the launch post

Released 12 November 2026 by iApp Technology, as open weights on Hugging Face.

What it does​

  • Writes while people speakText arrives 80, 320, 560 or 1120 ms behind the voice. You choose when you decode; the same weights serve every setting, and a longer wait is a little more accurate.
  • Any lengthThe included server cuts long recordings at pauses and decodes each piece on its own, as iApp's hosted service does, so a recording of an hour or more decodes in constant memory. We tested files of up to 87 minutes.
  • Thai script, or English words in EnglishWith the th-TH prompt everything comes out in Thai script. With th-mix, English words a speaker says come out in Latin letters.
  • Who said whatThe included server can label speakers in a meeting, with NVIDIA's open diarization model.
  • GPU or CPUOne NVIDIA H100 decodes files about 257 times faster than real time and keeps up with about 20 live streams. A laptop-class CPU with 8 threads keeps up with a live stream.
  • The model behind OpenThai ASR RealtimeiApp's hosted live speech-to-text service runs this model; see OpenThai ASR Realtime.

Measured​

Character error rate in %, lower is better, at 560 ms. We decoded every model ourselves with the same NeMo streaming script and the same settings: Typhoon ASR Streaming 0.6B as released, and with the 4-gram language model its authors publish for it.

Test setClipsOpenThai 2.0 ASR StreamingTyphoon ASR Streaming 0.6BTyphoon ASR Streaming 0.6B + its 4-gram LMDifference, ours − Typhoon
FLEURS, read speech1,0218.4711.2811.73−2.82 [−3.15, −2.48], ours lower
GigaSpeech2 test, YouTube8,82315.3516.3316.62−0.98 [−1.31, −0.58], ours lower
TVSpeech, TV and YouTube talk57012.4110.8411.05+1.57 [+0.90, +2.52], Typhoon lower
LOTUSDIS meetings, conference microphone5,51620.7728.9129.02−8.15 [−8.88, −7.46], ours lower
Isan dialect50034.9034.1535.79+0.75 [−0.21, +1.69], tie
  • Both sides of every comparison are written the same way before scoring: numbers made equal, spaces and punctuation removed. Differences are ours minus Typhoon ASR Streaming 0.6B, with a 95% paired bootstrap interval that resamples whole videos, recordings, speakers or meetings; an interval that includes zero is a tie. Bold marks an error rate lower than the other model's with 95% confidence.
  • TVSpeech is Typhoon ASR Streaming 0.6B's own benchmark. LOTUSDIS: we trained on other meetings from the same rooms and microphones, and some speakers appear in both, so that row favours us. FLEURS: we trained on its training and development splits, not its test split. Both models trained on the GigaSpeech2 training split. We did not train on Isan, and Typhoon ASR Streaming 0.6B's published training data lists none.
  • Our released file carries a blank offset of −1.0 chosen on development data. Given the same treatment, Typhoon ASR Streaming 0.6B's best offset on the same development data is -2.0, and its error becomes FLEURS, read speech 11.63, GigaSpeech2 test, YouTube 15.26, TVSpeech, TV and YouTube talk 10.68, LOTUSDIS meetings, conference microphone 26.71, Isan dialect 33.87.

In short, at 560 ms this model makes fewer errors than Typhoon ASR Streaming 0.6B on read speech (FLEURS 8.47% against 11.28%), YouTube clips (GigaSpeech2 15.35% against 16.33%) and meetings (LOTUSDIS, a set that favours us). It makes more on TV talk shows (TVSpeech 12.41% against 10.84%), and the two tie on the Isan dialect. Decode whole recordings in pieces cut at pauses, as the included server does; the "Whole recordings" tab shows why.

Run it​

Install NVIDIA NeMo (Speech) as the model card describes, then decode a list of 16 kHz mono files with NeMo's streaming script. This is how every number on this page was measured.

hf download iapp/OpenThai2.0-ASR-Streaming --local-dir OpenThai2.0-ASR-Streaming
echo '{"audio_filepath": "a.wav", "duration": 8.8, "text": ""}' > in.jsonl
python <NeMo Speech>/examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py \
  model_path=OpenThai2.0-ASR-Streaming/openthai2.0-asr-streaming_bias-1.0.nemo \
  dataset_manifest=in.jsonl output_path=out att_context_size="[56,6]" \
  target_lang=th-TH strip_lang_tags=true rnnt_decoding.greedy.max_symbols_per_step=2

att_context_size sets the latency: [56,0] 80 ms, [56,3] 320 ms, [56,6] 560 ms, [56,13] 1120 ms. target_lang=th-mix writes English words in Latin letters.

Tested with NeMo Speech cf724ac, PyTorch 2.13 and Python 3.12.

Prefer a hosted service? OpenThai ASR Realtime runs this model behind a WebSocket API, and Chinda ASR Pro v2 is iApp's more accurate service for recordings.

Limits​

  • Long recordings. Decoded as one stream, it drops words after quiet pauses. TVSpeech's videos, rebuilt with one second of silence between clips, score 23.54% that way and 12.66% cut at pauses. The included server cuts at pauses for you.
  • TV and YouTube talk. Typhoon ASR Streaming 0.6B is more accurate on TVSpeech. Part of the gap is hesitation sounds such as เอ่อ, which TVSpeech's transcripts keep and this model leaves out; most of the rest is English words.
  • English words. With th-TH they come out in Thai script. With th-mix, 13.7% of TVSpeech's English words come out in Latin letters, against 16.1% for Typhoon ASR Streaming 0.6B.
  • Dialects. Central Thai only. On Isan it gets 34.9% of characters wrong, about the same as Typhoon ASR Streaming 0.6B.
  • Distance from the microphone. On LOTUSDIS meetings at 1120 ms, the conference microphone gives 20.3% errors and the far channels bt3m and bt10m 56.5% and 53.5%.
  • Not measured: telephone audio (8 kHz). Not included: a custom vocabulary for names.

Training​

12,373 hours of Thai speech and 961 hours of English. 10,400 of the Thai hours are YouTube audio from the GigaSpeech2 training split, with its automatic transcripts; the rest is human-transcribed: the Thai Dialect Corpus (Central Thai only), iApp's own recordings, Common Voice, WanJuan-SiLu, THAI SER, LOTUSDIS, the Som TTS dataset, NECTEC's LOTUS and TSync, Thai Elderly Speech, Gowajee and the FLEURS training splits. NVIDIA's tokenizer was extended with Thai pieces and the decoder was trained afresh on it, starting from NVIDIA's encoder. 115,000 optimizer steps saw about 30,700 hours of audio. Every setting, including the blank offset stored in the released file, was chosen on development sets; no clip of any test set above was trained on. The sources, hours and recipe are in the model card.

Licence​

OpenMDW License Agreement, version 1.1, the licence of its base model. OpenThai 2.0 ASR Streaming is a Thai fine-tune of NVIDIA Nemotron 3.5 ASR Streaming 0.6B; NVIDIA's licence and original model card are included unchanged in the repository.

Built on NVIDIA Nemotron.

Built by iApp Technology. Trained and evaluated on NVIDIA H100 GPUs provided by Siam AI Corporation.

Siam AI Corporation