Licence & attribution

Served model: a third-party foundation model, Apache-2.0.

Architecture: a Whisper audio encoder (80 mel, 24 layers, d_model 1024, 16 heads, FFN 4096) with a Qwen3 text decoder (28 layers, GQA 16/8); long-context design.

This project's contribution: the Vietnamese adapter (LoRA r=64, alpha=128 — see moss_adapter/) and the surrounding service: API, queue, diarization, casing/punctuation restoration.

The repository also holds an older line based on Qwen/Qwen3-ASR-1.7B (50_cleaned_final_ckpt/, merged_ckpt/). It is not what this demo serves, and the numbers in docs/BENCHMARKS.md belong to that line — not to the deployed one.

The adapter was fine-tuned on a corpus assembled from several sources, including research-only ones: capleaf/viVoice (CC-BY-NC-SA), hustep-lab/VoxVietnam-Dataset (CC-BY-NC-4.0), linhtran92/viet_bud500 (adjudicated CC-BY-NC-SA), tensorxt/ViMedCSS (non-commercial) and doof-ferb/LSVSC (NC paper terms).

Consequence: the adapter weights and any merged checkpoint are NOT cleared for redistribution or commercial use. Every source corpus needs a licence review, or a retrain without the non-commercial sets, before publication.

This is a known limitation, not a statement that reuse is permitted.

The sample audio clips on the home page are third-party material. Each clip shows its source, licence and transcript source directly under the player, as licences such as CC BY require.

Keep that attribution intact if you reuse the clips.

SotaSpeech is the name of this project. It is not the name of the foundation model, and must not be presented as the foundation-model vendor's or Alibaba Qwen's original model.

Other product names mentioned in benchmark material belong to their respective owners.

The project's source code is released under Apache-2.0; see the LICENSE and NOTICE files in the repository.

The model weights are a separate question, governed by section 2 above.