LeKa model β€” OpenThai2.0 ASR Streaming (ONNX)

Thai streaming speech-to-text model files used by the LeKa desktop app. The app downloads them on first run and runs them locally with ONNX Runtime; no audio leaves the computer.

This is an ONNX export of iapp/OpenThai2.0-ASR-Streaming by iApp, a Thai fine-tune of NVIDIA nemotron-3.5-asr-streaming-0.6b. The weights are unchanged (FP32); only the file format differs.

Files

file what it is
encoder.onnx + encoder.onnx.data cache-aware FastConformer encoder, one streaming step; language prompt and joint encoder projection included
decoder.onnx RNN-T prediction network (2-layer LSTM) + joint prediction projection
joint.onnx joint network β†’ logits (14,988 tokens + blank)
frontend.json, mel_fb.bin, window.bin exact log-mel front-end constants (128 mels, n_fft 512, hop 160, pre-emphasis 0.97)
model.json streaming constants (chunk/shift sizes, cache shapes, prompt ids, blank id)
tokens.json vocabulary

Latency setting: 560 ms (att_context_size [56, 6]). Greedy decoding, max 2 symbols per frame, blank offset applied at decode time (βˆ’2.0 conversation, βˆ’1.0 standard).

Checked against the original NeMo decoder: identical text on all test clips.

Licence and credit

Released under the OpenMDW License Agreement, version 1.1 (see LICENSE), the same licence as the original model. Model by iApp Technology (OpenThai), built on NVIDIA Nemotron. All credit for training goes to them; this repository only repackages the model for ONNX Runtime.

Speaker labels

speaker.onnx + speaker_frontend.json, speaker_mel_fb.bin, speaker_window.bin: ONNX export of NVIDIA TitaNet-Large (speaker embeddings, 192-d), licensed CC-BY-4.0 by NVIDIA. LeKa uses it to tell speakers apart in a room meeting. Weights unchanged.

Speaker split (diarization)

diar.onnx + diar.json, diar_preenc.bin, diar_sil_emb.bin, diar_frontend.json, diar_mel_fb.bin, diar_window.bin: ONNX export of NVIDIA Nemotron-3-Diarization (streaming Sortformer, up to 8 speakers), licensed under the OpenMDW License Agreement, version 1.1 (see LICENSE) by NVIDIA.

file what it is
diar.onnx encoder + transformer + speaker heads over [speaker cache, FIFO, chunk]; weights quantized to INT8 (dynamic, MatMul) for CPU speed
diar_preenc.bin the 8-frame stacking projection (pre-encoder), float32 (512 Γ— 1024)
diar_sil_emb.bin the model's learned silence embedding
diar.json streaming settings used by LeKa (chunk 25 frames + 4 right context = 2.32 s latency, cache 264, FIFO 264) and speaker-cache constants
diar_frontend.json, diar_mel_fb.bin, diar_window.bin log-mel front-end constants of the diarizer

FlexAttention in the original was replaced by the equivalent plain attention for export (identical output); the speaker-cache logic runs in the app. Checked against NeMo's own streaming run: same speaker decisions (FP32), 99.9 % (INT8). LeKa uses it to tell speakers apart in a room meeting; TitaNet above remains as a fallback.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for chaiwatt/LeKa-model

Quantized
(4)
this model