Skip to content
Other

Speakilo / OneVoice On-Device Model Set

a local voice translator for Snapdragon 8 Gen 2

2 min readupdated

A ready-to-use set of models for offline (on-device) speech recognition, translation, and synthesis right on your smartphone with no cloud calls. Built by Vantuk for Qualcomm Snapdragon 8 Gen 2 (SM8550) mobile processors, the pipeline splits the workload between the NPU (Hexagon HTP v73 via QNN SDK) and the CPU (via Sherpa-ONNX and llama.cpp). Unlike standalone models, this bundle packs the entire voice processing loop for English, Korean, and Vietnamese into a single 1.877 GB archive. It is built for Android developers creating offline translators and voice assistants.

Key features

  • Full 5-stage pipeline: spoken language identification (Spoken LID powered by ecapa-lid), automatic speech recognition (ASR), source readiness check, machine translation, and text-to-speech (TTS).
  • Hybrid hardware execution: heavy stages like language ID and voice decoding are offloaded to the NPU (QNN HTP v73), while quantized ASR models and the LLM translator run on the CPU.
  • Multilingual stack (EN, KO, VI): includes SenseVoice INT8 for English and Korean, streaming Korean ASR, gipformer INT8 for Vietnamese, and the Hy-MT2-1.8B-Q4_0-enviko32k translation model.
  • Local voice synthesis: uses Piper family models (piper-en-lessac, piper-ko-kss, piper-vi-vais1000) with eSpeak-NG support and CPU fallback.

What you can do

  • Embed a real-time offline dialogue translator into an Android app for Snapdragon 8 Gen 2 devices.
  • Reuse individual converted modules from the pack (such as QNN builds of Piper TTS or ecapa-lid) to accelerate your own voice bots on a mobile NPU.
  • Benchmark llama.cpp + KleidiAI (with +dotprod+i8mm instructions) alongside Sherpa-ONNX on mobile ARM processors.

Example

A typical pipeline run from the model card: incoming audio first passes through ecapa-lid on the NPU to detect the language, gets transcribed via SenseVoice or gipformer on the CPU, is translated by the GGUF-formatted Hy-MT2-1.8B model, and is finally voiced in the target language by the hybrid Piper synthesizer.

Read moreDetails
Target hardware: Qualcomm Hexagon HTP v73 (soc_model = 43, htp_arch = 73, QNN SDK 2.51.0).
Repository contents: 388 files totaling 1.877 GB.
Inference stack: QNN HTP v73, Sherpa-ONNX, llama.cpp with KleidiAI optimizations (+dotprod+i8mm), and eSpeak-NG for phonemization.
Discuss Telegram 0 41

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # Vantuk/model_set_v73
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download Vantuk/model_set_v73 --local-dir ~/models/model_set_v73
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to model_set_v73
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of model_set_v73 (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

Vantuk/model_set_v73
Author
Vantuk
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

410

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments