Funyi

Funyi is a local speech-to-text and live captions app built on Qwen/Qwen3-ASR-1.7B. It runs on your machine, starts a local realtime ASR WebSocket service, and provides a lightweight Tauri desktop caption client.

Funyi is designed for local, single-user use. It is not a hosted service or a public multi-user ASR server.

Demo

funyi-desktop-demo.mp4

Requirements

• Python 3.10 or newer; Python 3.12 is recommended.
• uv for Python dependencies.
• Linux or WSL with an NVIDIA CUDA GPU for the realtime backend.
• Access to download the ASR, forced-aligner, and translation models, or local model directories.
• For the desktop client: Node.js with Corepack-enabled pnpm, Rust/Cargo, and Windows or macOS native build tools.

Windows desktop builds also need Visual Studio Build Tools 2022 with the Desktop development with C++ workload and a Windows 10/11 SDK.

The desktop client currently runs on Windows and macOS.

Quick Start

1. Start The Backend

Run these commands from the repository root in Linux or WSL.

Install Python dependencies:

uv sync --python 3.12 --frozen

Start the backend in one terminal. On a fresh checkout or empty model cache, use the download target once. The first start can take a while because it downloads and warms the models:

FUNYI_ALLOW_DOWNLOADS=1 ./scripts/start_backend.sh

After the models are cached, use:

./scripts/start_backend.sh

Check the backend:

curl http://127.0.0.1:8000/healthz

2. Start The Desktop Client

Run the desktop client from a Windows or macOS checkout. On Windows, use a Windows checkout so system-audio capture uses native WASAPI.

Install desktop dependencies once:

cd desktop
corepack pnpm install

Start the desktop client:

corepack pnpm run dev

In the desktop UI, connect to:

ws://127.0.0.1:8000/ws/asr

Choose your audio source, optionally choose a speech language and translation target, then start captions.

Windows Desktop

Run the backend in Linux or WSL. Start the desktop client from a Windows checkout using the desktop commands above, then connect it to the backend URL, usually ws://127.0.0.1:8000/ws/asr.

Common Runs

Command	What it does
FUNYI_ALLOW_DOWNLOADS=1 ./scripts/start_backend.sh	Start the full backend and allow forced-aligner/translation model downloads.
./scripts/start_backend.sh	Start the full backend from cached or local models.
FUNYI_TRANSLATION_MODEL= ./scripts/start_backend.sh	Start ASR plus the required forced aligner, without translation.
FUNYI_PORT=8001 ./scripts/start_backend.sh	Start the backend on another port.
./scripts/start_backend.sh --no-vad	Start without VAD speech gating; all received audio is passed to ASR.
corepack pnpm install	Install desktop dependencies from desktop/.
corepack pnpm run dev	Start the desktop client from desktop/.

To use local model directories instead of Hugging Face model ids:

FUNYI_ASR_MODEL=/path/to/Qwen3-ASR-1.7B \
FUNYI_TIMESTAMP_MODEL=/path/to/Qwen3-ForcedAligner-0.6B \
FUNYI_TRANSLATION_MODEL=/path/to/Hy-MT2-1.8B \
./scripts/start_backend.sh

Realtime ASR requires the forced aligner. Translation is available when the backend starts with a translation model and the client requests a target language.

Privacy

Audio sent to ws://127.0.0.1:8000/ws/asr is processed by the local Python service.

Troubleshooting

• If downloads fail, run FUNYI_ALLOW_DOWNLOADS=1 ./scripts/start_backend.sh once or pass local model paths.
• If the backend fails to start, check CUDA availability and model paths.
• If the desktop cannot connect, confirm the URL is ws://127.0.0.1:8000/ws/asr and curl http://127.0.0.1:8000/healthz returns {"status":"ok"}.
• If pnpm is not found, run desktop commands as corepack pnpm ....
• On Windows, run the desktop client from a Windows checkout when validating WASAPI loopback. WSL is not the right place to validate Windows system-audio capture.

Documentation

• desktop/README.md: desktop client details and OS audio-capture notes.
• docs/realtime_asr_service.md: WebSocket protocol, timestamp behavior, and realtime service rules.
• docs/realtime_translation_design.md: translation behavior and target-language details.

License

Funyi's original project code is licensed under the MIT License. See LICENSE.

The files under qwen3_asr_runtime/hf_qwen3_asr/ are copied from Qwen3-ASR / Hugging Face Transformers integration code and retain their upstream Apache-2.0 notices. See THIRD_PARTY_NOTICES.md and LICENSES/Apache-2.0.txt.

Third-party dependencies, model weights, and external datasets remain under their own license terms.