typhoon-asr-streaming-115m

Typhoon ASR Streaming 115M

Cache-aware streaming Thai ASR (FastConformer-Transducer, 115M params). Converted from Typhoon ASR Real-time — causal convolutions + chunked limited-context attention swapped in, every weight copied as a warm start, then briefly fine-tuned on the ~11k-hour Thai corpus. Unlike the full-context base (which collapses to 62.8% CER when forced to stream at 1040 ms), this model decodes live, chunk by chunk, with the encoder cache carried across chunks.

At a glance

Use this model for low latency / low cost when English code-switching is light; for best accuracy use typhoon-ai/typhoon-asr-streaming-nemotron-0.6b.

Usage

Requires NVIDIA NeMo with cache-aware streaming support (see the project repo for the pinned commit and the full streaming loop / server / demo):

import nemo.collections.asr as nemo_asr

model = nemo_asr.models.ASRModel.restore_from("typhoon-asr-streaming-115m.nemo",
                                              map_location="cuda")
model.eval()

# offline (whole file)
text = model.transcribe(audio=["clip.wav"])[0].text

# true streaming: model.encoder.setup_streaming_params(chunk_size=13, left_chunks=2,
# shift_size=13) + conformer_stream_step(...) — see demo/asr_engine.py in the repo.

Decode-time vocabulary steering (phrase boosting + n-gram shallow fusion) runs inside the streaming decoder — see the repo's docs/SHALLOW_FUSION.md.

License & attribution

Released under CC-BY-4.0, following the base-model chain:

• fine-tuned + converted from scb10x/typhoon-asr-realtime (CC-BY-4.0, SCB DataX),
• itself fine-tuned from nvidia/stt_en_fastconformer_transducer_large (CC-BY-4.0, NVIDIA).

Modifications in this release: full-context → cache-aware streaming conversion (causal convolutions, chunked limited-context attention, multi-latency att_context_size) and a 1-epoch Thai fine-tune on the Typhoon ASR Real-time corpus.