An on-device intelligent voice assistant designed to boost creation and recording efficiency. Built with Tauri 2 (Rust backend) + Vite (TypeScript frontend), it delivers smooth realtime STT (SenseVoice ONNX / FunASR local inference) and powerful TTS (Rust libtorch + Python bridge + local/cloud models) while running entirely on your local device to protect privacy.
- Core capabilities: on-device realtime recording and recognition (cursor-follow injection supported), realtime translation, AI proofreading and summarization, text-to-speech and voice cloning
- Main stack: Tauri 2 (Rust), Vite (TypeScript), ONNX Runtime (Rust
ort), Python (PyTorch/Transformers) - Scenarios: meeting/interview notes, bilingual captions, speak-while-you-type (input injection), dubbing and voice cloning
- SenseVoice ONNX and FunASR local inference with realtime and final results
- Upgraded
native-sttpath for offline audio-file transcription with Whisper / SenseVoice native backends - Three output modes: cursor-follow injection (recommended) / realtime output / final auto-enter
- Global hotkey
RightAltto start/stop recording anytime - Supported languages:
auto/zh/en/yue/ja/ko/nospeech - Mini Mode: floating realtime captions window at the bottom-right
Download STT model (SenseVoice ONNX): SenseVoiceSmall
modelscope download --model iic/SenseVoiceSmall --local_dir ./SenseVoiceSmall
Download STT model (FunASR):
modelscope download --model FunAudioLLM/Fun-ASR-Nano-2512 --local_dir ./Fun-ASR-Nano-2512
Speaker Diarization (optional, sherpa-onnx) — required if you want to distinguish Speaker_1 / Speaker_2 / Speaker_3 in real-time STT. The diarization runtime is driven by sherpa-onnx and needs two extra ONNX files alongside the main STT model:
segmentation.onnx— recommended:sherpa-onnx-pyannote-segmentation-3-0embedding.onnx— recommended:3dspeaker_speech_eres2net_base_sv_zh-cn_3dspeaker_16k.onnx
Download from the sherpa-onnx GitHub Releases:
- Segmentation models: https://github.com/k2-fsa/sherpa-onnx/releases/tag/speaker-segmentation-models
- Embedding models: https://github.com/k2-fsa/sherpa-onnx/releases/tag/speaker-recongition-models
- Official model guide: https://k2-fsa.github.io/sherpa/onnx/speaker-diarization/models.html
Place both files in a folder (e.g. ./speaker-diarization/) and point Speaker Segmentation Model / Speaker Embedding Model in settings to the respective .onnx files (or point both fields at the same folder — the app auto-detects them). See docs/html/appendix-en.html for the full beginner walkthrough.
- Built-in translation proxy with Ollama / OpenAI-compatible APIs
- Recommended local translation model:
ZimaBlueAI/HY-MT1.5-1.8(Ollama) - Bilingual captions in the same view with streaming output
- Strategies: translate only final results (saves API quota) or translate realtime results too
- AI proofreading ("Correction" feature): auto-polish recognized text with configurable concurrency (default 2, max 4)
- Smart summaries: scheduled meeting/recording summaries (recommended every 60 seconds), shown as a tree or queue
- Supports OpenAI / Ollama / Volcengine; recommended local model:
qwen3:4b
- Live mode: while recording, the main "LingLu" output panel grows a live topic tree in the dock area; closed topics never get rewritten, the active topic is highlighted with a vermilion left border; detach to a floating window for second-screen use
- Wrap-up: after stop, auto-consolidates summary / key decisions / action items, with scenario-aware field weighting (personal / gathering never force-extracts action items)
- Delivery: pick from 5 HTML report themes (terminal-grade / keynote-stage / editorial / minimal-print / sumi-ink Chinese aesthetic); backend-controlled rendering + CSP + XSS defense
- 9 LLM providers: Ollama / OpenAI / DeepSeek / Kimi / GLM / Volcengine / MiniMax / Anthropic Claude / Google Gemini — all accept third-party OpenAI-compatible proxy URLs
- 8 scenario templates: generic / personal / private chat / gathering / closed / public / business / academic; with scenario=auto, confidence < 0.5 forcibly falls back to generic
- On-device first: Ollama (local) is the default — no API key, zero per-call cost, data stays on your machine
- Session-anchored: after stop, artifacts land in
session/<id>/linglu.jsonandsession/<id>/reports/<theme>.html, persisting across restarts - See docs/html/linglu-en.html and docs/v0.4-linglu/ for details
- Qwen3-TTS with three modes:
Base: high-quality synthesis without reference audioCustomVoice: clone a voice with 5-15 seconds of reference audioVoiceDesign: generate a new voice from text prompts (no reference audio)
- Index-TTS2 emotional vector control + prompt guidance for more expressive cloning
- Cloud API support:
- Volcengine: rich high-quality voices, supports dialects/foreign languages
- OpenAI TTS: alloy / echo / fable / onyx / nova / shimmer
- MiniMax: high-expressiveness TTS models
- TTS runtime can be packaged in the full installer or downloaded on demand via Lite packages
- Dedicated
STS Speech Workbench— does not reuse the mainSTT / TTSsettings - Configure the workbench first, then click
Open Mini Widget - The
Miniwidget supports starting a recording by clicking it or pressingRightAlt - Click again or press
RightAltagain to stop recording - Translation, voice cloning, and auto-playback run only after recording stops
- Clicking again while processing cancels the current task and restarts recording
- System tray: close the window to minimize to tray, restore or exit from the tray menu
- Download and install the latest Release package, then double-click to launch
- Open settings, select STT backend (
onnxorfunasr), and point Model Directory to the corresponding model folder - Press
RightAltto start recording
If you want offline file transcription instead of live microphone STT, switch the STT backend to Whisper or SenseVoice, then select a model file and an audio file in the STT panel.
For model downloads and FAQs, see INSTALL.md.
cd /path/to/boundless-flow
pnpm install
pnpm run tauri:devFull environment setup (Rust/MSVC, Python/TTS, Lite packages, packaging outputs) is in INSTALL.md.
| Setting | Description | Recommendation |
|---|---|---|
| Model Directory | ONNX backend requires model.onnx and tokens.json; FunASR backend should point to the full Fun-ASR-Nano-2512 model directory |
Point to the exact model directory |
| Backend | onnx / funasr for realtime mic STT, whisper / sensevoice for native offline file transcription |
Use onnx or funasr for live subtitles; switch only when transcribing files |
| Frame Interval (ms) | Audio frame send frequency; lower is more realtime but higher CPU | 20ms |
| Language | auto/zh/en/yue/ja/ko/nospeech | auto |
| TextNorm | Text normalization | auto |
| Output Mode | Cursor-follow injection / realtime output / final auto-enter | Choose by scenario |
Auto-downgrades if a platform does not support a feature (e.g., cursor-follow injection may be unavailable on some platforms).
native-sttis currently designed for uploaded audio files, not live microphone subtitles; for realtime mic subtitles useonnxorfunasr.
| Setting | Example |
|---|---|
| Translation API Base URL | Ollama: http://localhost:11434 (no /v1; auto-calls /api/chat) — /v1 is also accepted and stripped internally. OpenAI-compatible: https://api.openai.com/v1 |
| Translation Model | Ollama models need the :tag suffix, e.g. ZimaBlueAI/HY-MT1.5-1.8:1.8b; OpenAI-compat: gpt-4o-mini / translategemma |
| Translation API Key | Can be empty for Ollama |
| Streaming Output | Recommended (smoother for local models) |
| Goal | Recommended combination |
|---|---|
| Fastest end-to-end pipeline | Local onnx/funasr + OpenAI-compatible translation + volcengine_tts |
| Validate voice cloning first | Local onnx/funasr + OpenAI-compatible translation + qwen3_tts/index_tts2 + reference audio |
Recommended order:
- Run the "fastest pipeline" first to confirm recording, translation, and playback all work
- Switch to "voice-cloning-first" to validate reference audio and reference text
- Finally swap in your target model combination
For full STS Mini field guidance, verification steps, and troubleshooting see:
| Setting | Description | Example |
|---|---|---|
| Enable proofreading & summarization (LLM) | Global toggle | Enabled |
| Summary Provider | OpenAI / Ollama / Volcengine | Ollama |
| Summary API Base URL | Service endpoint (Ollama: no /v1) |
http://localhost:11434 |
| Summary Model | Model ID | qwen3:4b / doubao-seed-1-6 |
| Proofreading Concurrency | 1-4 | 2 |
| Summary Update Interval (s) | Frequency of summaries | 60 |
Entry: Main panel → Advanced → LingLu API · Live topic tree + reports (open by default). Or fill it in once during onboarding step 4.
| Setting | Description | Recommended |
|---|---|---|
| Provider | 9 options; Ollama listed first as "Local (recommended)" | ollama |
| Base URL | Any OpenAI-compatible proxy works | http://localhost:11434 or http://192.168.x.x:11434 |
| API Key | Leave empty for Ollama; required for others | — |
| Model | Model id (include tag) | qwen3.6:35b / deepseek-chat / claude-3-5-haiku-20241022 |
| Scenario | auto + 8 scenarios; auto with confidence < 0.5 falls back to generic | auto |
| Report Theme | 5 themes; smart default per scenario | sumi-ink (personal/gathering) / terminal-grade (business) |
| Temperature | 0.0–1.0; 0.1–0.3 for minutes | 0.2 |
| Max tokens | ≥ 2048 for reasoning models (chain-of-thought consumes budget) | 4096 |
| Interval (seconds) | Lower bound 10, default 60 | 60 |
Full prompts and 8-scenario templates: docs/html/linglu-en.html. Design docs (requirements / architecture / implementation / audit): docs/v0.4-linglu/.
Qwen3-TTS:
modelscope download --model Qwen/Qwen3-TTS-12Hz-1.7B-Base --local_dir ./Qwen/Qwen3-TTS-12Hz-1.7B-Base
modelscope download --model Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --local_dir ./Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
modelscope download --model Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign --local_dir ./Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesignIndex-TTS2:
modelscope download --model IndexTeam/IndexTTS-2 --local_dir ./IndexTeam/IndexTTS-2After downloading, set TTS Model Directory to the corresponding model folder in settings.
In settings, choose Volcengine TTS and fill in:
| Field | Description |
|---|---|
| AppId | Volcengine app identifier |
| Token | Access token |
| Cluster | Cluster identifier (e.g., volcano_tts) |
| VoiceType | Voice identifier |
Optional: UID, audio encoding, sample rate, speed/volume/pitch multipliers, emotion, and more.
The Boundless Flow frontend calls backend commands via Tauri invoke, including:
- Config:
get_app_config/set_app_config - Recording:
start_listening/stop_listening - Injection:
inject_text - Translation proxy:
translate_via_backend - TTS:
tts_generate/tts_read_audio_base64
Backend entrypoint: src-tauri/src/main.rs.
- Frontend:
src/(Vite) - Rust backend:
src-tauri/ - Python bridge (TTS):
src-tauri/python/ - Tauri config:
src-tauri/tauri.conf.json(plus platform-specific configs) - Docs:
docs/(detailed user guides in Chinese and English)
docs/index.html: documentation entry and core capabilitiesdocs/stt.html: realtime STT and model configurationdocs/translation.html: realtime translation flow and settingsdocs/proofreading-summary.html: AI proofreading and smart summariesdocs/tts-voice-cloning.html: speech synthesis and voice cloningdocs/appendix.html: model downloads, sherpa-onnx speaker diarization, and API configuration guidedocs/context-landing.html: design philosophy, quick start, and best practices landing page
English docs are located alongside as *-en.html.
Local checks before submitting:
cd /path/to/boundless-flow
pnpm run type-check
pnpm run build
pnpm run tauri:build
pnpm run tauri:bundleRecommended submission: a single goal, reproducible steps, screenshots/logs (especially for UI/audio issues).
Copyright 2026 ZimaBlueAI & WaytoAGI-dev. All rights reserved.