typhoon-ai/typhoon-asr-streaming-nemotron-0.6b-ngram

Typhoon ASR Streaming Nemotron 0.6B — paired 4-gram fusion LM

4-gram over sub-word tokens (aggregate 15,135 vocabulary) for decode-time shallow fusion with typhoon-ai/typhoon-asr-streaming-nemotron-0.6b. Trained on the ASR training transcripts plus a synthetic in-domain corpus (English code-switch terms inserted into LLM-generated Thai carrier templates) — no evaluation data. Tokenizer-locked: it works only with the matching model's tokenizer.

File: ngram_4gram_agg.nemo (~2.4 GB, NeMo NGramGPULanguageModel binary). It ships as a .nemo binary rather than ARPA because the 15k aggregate vocabulary corrupts the ARPA chr() token encoding — the binary is offset-independent and is the supported artifact.

Usage

Enable n-gram fusion in the streaming greedy decoder (see the project repo, docs/SHALLOW_FUSION.md):

from omegaconf import OmegaConf

cfg = {"strategy": "greedy_batch", "greedy": {
    "ngram_lm_model": "ngram_4gram_agg.nemo",
    "ngram_lm_alpha": 0.5,
}}
model.change_decoding_strategy(OmegaConf.create(cfg))
model.set_inference_prompt("th-TH")   # re-apply after changing decoding strategy
model.decoding.set_strip_lang_tags(True)

On the TVSpeech held-out test @1040 ms this lifts keyword recall 16.6% → 20.4% at neutral CER, with under a 3% change in the real-time factor.

License

OpenMDW-1.1 (see LICENSE in this repository), matching the paired acoustic model, which derives from NVIDIA's Nemotron streaming ASR base. "Nemotron" is an NVIDIA trademark, used only to identify the base model; this release is not affiliated with or endorsed by NVIDIA. Built by SCB DataX for the Typhoon ASR Streaming project.