---
license: other
license_name: openmdw-1.1
license_link: https://openmdw.ai/license/1-1/
language:
  - th
  - en
library_name: nemo
pipeline_tag: automatic-speech-recognition
base_model: nvidia/nemotron-3.5-asr-streaming-0.6b
tags:
  - automatic-speech-recognition
  - speech
  - streaming
  - FastConformer
  - Transducer
  - Thai
  - NeMo
---

# Typhoon ASR Streaming Nemotron 0.6B

Cache-aware **streaming** Thai ASR (FastConformer-Transducer, 0.6B params,
prompt-conditioned) — the **best-accuracy model** of the Typhoon ASR Streaming release.
Adapted to Thai from NVIDIA's Nemotron streaming ASR base: the multilingual tokenizer is
**extended** (not replaced) with the multi-character Thai BPE from Typhoon ASR Real-time,
keeping the other languages recoverable, followed by Thai adaptation and a numeric
fine-tuning stage.


## At a glance

| | |
|---|---|
| Architecture | FastConformer-Transducer (RNN-T), cache-aware streaming, prompt-conditioned |
| NeMo class | `EncDecRNNTBPEModelWithPrompt` |
| Tokenizer | aggregate 15,135 = Nemotron multilingual 13,087 + Thai BPE 2,048 |
| Language prompt | **required**: `target_lang="th-TH"` |
| Latency knob | `chunk_size` × 80 ms → 1040 / 480 / 80 ms |
| Streaming CER | **14.1%** TVSpeech · **9.3%** GigaSpeech2-Thai @1040 ms (13.0% TVSpeech @3200 ms) |
| Steering (TVSpeech held-out, 1040 ms) | keyword recall 16.6% → **20.7%** (phrase boost) at no CER cost |
| RTF (batch-1, H100 @1040 ms) | 0.024 (~42× real time); batch-16 ≈ 300 concurrent streams/GPU |
| Paired n-gram | [`typhoon-ai/typhoon-asr-streaming-nemotron-0.6b-ngram`](https://huggingface.co/typhoon-ai/typhoon-asr-streaming-nemotron-0.6b-ngram) |

## Usage

Requires NVIDIA NeMo with cache-aware streaming and `EncDecRNNTBPEModelWithPrompt` support
(see the [project repo](https://github.com/warit-s/typhoon-asr-streaming) for the pinned
commit and the full streaming loop / server / demo):

```python
import nemo.collections.asr as nemo_asr

model = nemo_asr.models.ASRModel.restore_from(
    "typhoon-asr-streaming-nemotron-0.6b.nemo", map_location="cuda")
model.eval()
model.set_inference_prompt("th-TH")          # required (prompt-conditioned)
model.decoding.set_strip_lang_tags(True)     # strip the <xx-XX> tag from output

# offline (whole file)
text = model.transcribe(audio=["clip.wav"], target_lang="th-TH",
                        return_hypotheses=True)[0].text

# true streaming: setup_streaming_params + conformer_stream_step — see the repo.
```

Decode-time steering (phrase boosting / n-gram fusion) works inside the streaming decoder;
with this aggregate-tokenizer model set `boosting_tree.source_lang="th-TH"` and use the
paired `.nemo` n-gram — see the repo's `docs/SHALLOW_FUSION.md`.

## License & attribution

Released under **OpenMDW-1.1** (see [`LICENSE`](LICENSE) in this repository), following the
base model: **[NVIDIA Nemotron 3.5 ASR Streaming 0.6B](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b)** (OpenMDW-1.1). The base model's
copyright and origin notices are preserved in this repository as required by the license.

Modifications in this release: aggregate tokenizer extension with Thai BPE, Thai
adaptation on the ~11k-hour Typhoon corpus, and numeric fine-tuning.

> **Trademark note:** "Nemotron" is a trademark of NVIDIA Corporation. It is used here
> nominatively, solely to identify the base model this checkpoint derives from. This
> release is not affiliated with, sponsored by, or endorsed by NVIDIA.