LakoreAI/onevoice-vega

OneVoice / Vega — on-device speech translation model bundle

The full ONNX model bundle for QUAVIO/Vega, a 100% offline (no-cloud) speech-to-speech translator for noisy industrial environments, built for the Saigon AI Hub × Qualcomm "OneVoice AI Challenge". One cascaded engine (denoise → VAD → ASR → MT → TTS) shipped on the Qualcomm Dragonwing QCS6490 (Rubik Pi 3 / RB3 Gen 2), in two carriers:

• Vega Hand — ruggedized push-to-talk handheld (one-to-one).
• Vega Air — drone-mounted multilingual broadcast payload (one-to-many).

Languages: Vietnamese ↔ English / Chinese (Mandarin) / Korean (6 directions).

This repo holds the .onnx artifacts (and Piper's espeak-ng-data phoneme tables) that the Rust runtime (src/) loads via ONNX Runtime / QNN. It is the same tree scripts/pack_models.sh ships as the GitHub Release bundle and that scripts/setup_models.sh installs.

Models

Stage	Model	Size	License	Source
VAD	Silero VAD	2.1 MB	MIT	onnx-community/silero-vad
ASR (VI)	PhoWhisper-base (int8)	22 MB enc / 51 MB dec	BSD-3-Clause	huuquyet/PhoWhisper-base
ASR (EN/ZH/KO)	Whisper-small (int8)	88 MB enc / 149 MB dec	MIT	onnx-community/whisper-small
MT (VI↔EN)	Opus-MT (INT8)	45 MB enc / 78 MB dec	Apache-2.0 / CC-BY	Helsinki-NLP/opus-mt-*
TTS (VI)	Piper vais1000-medium (VITS)	60 MB	CC BY 4.0	rhasspy/piper-voices
TTS (EN)	Piper lessac-medium (VITS)	60 MB	CC BY 4.0	rhasspy/piper-voices
TTS (ZH)	Piper huayan-medium (VITS)	60 MB	CC BY 4.0	rhasspy/piper-voices
TTS (KO)	MMS-TTS Korean (VITS)	109 MB	MIT	facebook/mms-tts-kor
TTS runtime data	espeak-ng-data (Piper phonemes)	19 MB	GPL-3.0	piper-tts PyPI package

Denoise (RNNoise) has no model file — its weights are compiled into the Rust nnnoiseless crate. Encoders/decoders use the merged-decoder KV-cache ONNX layout (decoder_model_merged.onnx) for autoregressive greedy decode.

Repository structure

├── vad/model.onnx
├── asr/vi/{encoder_model.onnx, decoder_model_merged.onnx, tokenizer.json, ...}
├── asr/en-zh-ko/{...}
├── mt/vi-en/{encoder_model.onnx, decoder_model_merged.onnx, tokenizer.json, ...}
├── mt/en-vi/{...}
├── tts/{vi,en,zh,ko}/*.onnx (+ Piper *.onnx.json sidecars)
└── tts/espeak-ng-data/       # Piper phoneme data (vi/en/zh)

Each directory carries a MANIFEST.md recording source checkpoint, export date, quantization scheme, and eval metrics.

Usage

git lfs install
git clone https://huggingface.co/LakoreAI/onevoice-vega models/

# run the full Rust pipeline (denoise → VAD → ASR → MT → TTS)
# see the project repo: https://github.com/MinLee0210/onevoice-vega
scripts/run_pipeline.sh data/eval/ptt_sample_vi_48k.wav out.wav -- --source vi --target en

Or download only what setup_models.sh needs from the project repo — it points at this bundle as the "rebuild" source for vad/asr/tts.

Performance

Measured on the QCS6490 (Hexagon v68 HTP), full-INT8 QNN context binaries — see the companion qualcomm-ai-hub-community/whisper-*-qcs6490-qnn-int8 repos:

Stage	Latency
ASR encoder (PhoWhisper-base, 30 s)	~49.5 ms (NPU)
ASR decoder (per token)	~12.1 ms (NPU)
End-to-end pipeline (CPU dev machine)	~741 ms button-release → audio

License note

This is a mixed-license bundle — each stage carries its own license (table above). In particular, espeak-ng-data/ is GPL-3.0 (sourced from the piper-tts package); embedding it in a distributed binary raises an unresolved GPL question, relevant only to commercial distribution, not to the prototype/demo. See docs/decisions/piper-vi-voice-license-trap.md in the project repo for the full analysis. The Korean TTS path (MMS-TTS, MIT) avoids espeak entirely.

Attribution

• Whisper — OpenAI (MIT); PhoWhisper — VinAI / huuquyet.
• Opus-MT — Helsinki-NLP (Apache-2.0).
• Silero VAD — Silero Team (MIT).
• Piper voices — rhasspy/piper-voices; espeak-ng data from piper-tts.
• MMS-TTS — Meta (MIT).

Project: MinLee0210/onevoice-vega.