Typhoon ASR Streaming 115M — paired 4-gram fusion LM
KenLM 4-gram over sub-word tokens (2048 Thai BPE) for decode-time shallow fusion with
typhoon-ai/typhoon-asr-streaming-115m.
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_bpe2048.arpa (~2.7 GB, KenLM ARPA).
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_bpe2048.arpa",
"ngram_lm_alpha": 0.5,
}}
model.change_decoding_strategy(OmegaConf.create(cfg))
Fusion re-ranks candidates inside the streaming decoder and changes the real-time factor by under 3%. Typical weight range: 0.3 (broad domain) – 0.7 (tight in-domain).
License
CC-BY-4.0, matching the paired acoustic model. Built by SCB DataX for the Typhoon ASR Streaming project.