Appendix (Beginner-friendly)

If this is your first time dealing with “model downloads / local LLMs / cloud APIs”, follow the steps below.

Appendix A: ModelScope (Model Downloads)

When you need it: Download local STT (SenseVoice) and local TTS (Qwen3-TTS / Index-TTS2) model files.

Official docs: https://www.modelscope.cn/docs/

1) Do I need to sign up?

Most public models can be downloaded without login. If you see an access/permission error, you typically need to log in on ModelScope and request access as prompted.

2) Check your OS / CPU architecture

• Windows: Settings → System → About → System type; or PowerShell $env:PROCESSOR_ARCHITECTURE
• macOS: Apple menu → About This Mac → Chip; or Terminal uname -m
• Linux: Terminal uname -m

3) Install ModelScope CLI

The download command comes from the ModelScope CLI (requires Python + pip):

pip install modelscope
modelscope --help

4) Find models (optional)

Use the ModelScope website search, open the model page, and copy the model ID.

5) Download examples (aligned with this manual)

modelscope download --model iic/SenseVoiceSmall --local_dir ./SenseVoiceSmall
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-VoiceDesign
modelscope download --model IndexTeam/IndexTTS-2 --local_dir ./IndexTeam/IndexTTS-2

6) Where do I set the path in Boundless Flow?

• STT Model Directory: point to the SenseVoice download folder
• TTS Model Directory: point to the specific TTS model folder you want to use

Appendix B: Ollama (Local LLMs for Translation / Summary)

When you need it: Run translation/summary models locally or on your LAN. Boundless Flow calls Ollama's native /api/chat endpoint, so the Base URL only needs the server address — do not append /v1 (typical Base URL: http://localhost:11434).

Official docs: https://docs.ollama.com/

API references: Ollama native API · OpenAI compatibility

1) Install

Follow the official “Download / Get started” instructions for your OS and CPU architecture.

2) Verify

ollama --version

3) Pull models (aligned with this manual)

ollama pull ZimaBlueAI/HY-MT1.5-1.8:1.8b   # or :7b
ollama pull qwen3:4b

Heads up: this translation model has no latest tag in Ollama, so you must include the :tag suffix (e.g. :1.8b or :7b). Without it Ollama returns model 'ZimaBlueAI/HY-MT1.5-1.8' not found.

4) Suggested settings in Boundless Flow

• Translation: Base URL http://localhost:11434 (local) or http://<LAN-IP>:11434 (LAN); Model ZimaBlueAI/HY-MT1.5-1.8:1.8b
• Proofreading/Summary: Base URL http://localhost:11434; Model qwen3:4b
• STS workbench → Translation API Base URL: same as above, http://localhost:11434

Compatibility note: the recommended form is without /v1 (which routes to Ollama's native /api/chat). If you paste a Base URL that ends in /v1, Boundless Flow will recognize the Ollama port and strip the suffix automatically — so all three places (realtime STT translation, summary, STS translation) behave identically.

Appendix C: Volcengine Cloud TTS

When you need it: Use Volcengine cloud voices / higher quality TTS / voice replication features.

Useful links:

• Volcengine TTS product page
• Speech docs hub
• VoiceType list

1) Do I need to sign up?

Yes. You need a Volcengine account and enabled speech services in the console.

2) What info do I need?

• AppId
• Token
• Cluster (e.g. volcano_tts / volcengine_tts, depends on your console)
• VoiceType (pick from the VoiceType list)

3) Minimal setup in Boundless Flow

1. Settings → TTS Model: select Volcengine TTS
2. Choose HTTP mode
3. Fill AppId / Token / Cluster / VoiceType
4. Synthesize a short sentence to verify audio output

Appendix D: sherpa-onnx Speaker Diarization

When you need it: Separate Speaker_1 / Speaker_2 / Speaker_3 in real-time STT. In addition to the main STT model (SenseVoice / FunASR), you need two extra ONNX files driven by the sherpa-onnx runtime.

Official repo: https://github.com/k2-fsa/sherpa-onnx

Official model guide: sherpa-onnx speaker diarization models

1) Two files you need

• segmentation.onnx — speaker segmentation model (recommended: sherpa-onnx-pyannote-segmentation-3-0)
• embedding.onnx — speaker embedding model (recommended: 3dspeaker_speech_eres2net_base_sv_zh-cn_3dspeaker_16k.onnx)

2) Where to download

• Segmentation models: speaker-segmentation-models
• Embedding models: speaker-recongition-models

Download the .onnx files directly from the GitHub Releases above — no need to build sherpa-onnx from source. If the asset is an archive, extract segmentation.onnx and embedding.onnx first and keep the filenames as-is.

3) Recommended directory layout

./speaker-diarization/
  segmentation.onnx
  embedding.onnx

4) Where do I set the paths in Boundless Flow?

• Speaker segmentation model: point to segmentation.onnx
• Speaker embedding model: point to embedding.onnx

You can also point both fields to the same directory — the app will auto-detect segmentation.onnx and embedding.onnx inside it.

5) Notes

• The SenseVoice / FunASR "Model Directory" only handles speech-to-text; it cannot replace the speaker diarization models.
• Speaker diarization in this project supports ONNX only. A Fun-ASR-Nano-2512 folder cannot substitute for segmentation.onnx / embedding.onnx.
• If diarization still fails, the UI now surfaces a readable error. The most common cause is an onnx file whose schema doesn't match the bundled sherpa-onnx runtime — start with the two recommended models above.