Real-time Speech-to-Text (STT) & Model Selection

The core feature of Boundless Flow is real-time Speech-to-Text (STT) based on local models. It accurately and rapidly converts your speech into text and supports multiple output methods.

Real-time STT Introduction

Boundless Flow supports both SenseVoice ONNX and FunASR for local real-time STT. Both paths provide interim streaming output and final sentence-level output, so text can appear while you are speaking.

Output Methods

In the settings, you can choose different output methods:

• Cursor-following Injection (Recommended): Speak as you write. The recognized text is directly injected into your current cursor position. After enabling this, please click where you want to type (e.g., Word, web input box) before you start speaking.
• Real-time Output: Displays the recognized text in real-time within the application interface.
• Final Auto-Return: Automatically outputs the final result and adds a line break after you finish a sentence.

How to Start Real-time STT

Selecting and Configuring Different Models

For the speech recognition feature to work, a simple model configuration is required:

Select Backend and Configure Model Directory

1. Open the main application interface and go to Settings.
2. Select the STT backend you want to use (ONNX or FunASR).
3. Locate Model Directory and point it to the corresponding model folder.

Recommended Model Download (ModelScope)

Boundless Flow uses a local SenseVoice ONNX model by default. Recommended download via ModelScope (see ModelScope docs; beginner steps: Appendix A):

modelscope download --model iic/SenseVoiceSmall --local_dir ./SenseVoiceSmall

After downloading, set Model Directory to the download folder (e.g., ./SenseVoiceSmall or an absolute Windows path).

If you choose the FunASR backend, download the FunASR-Nano model and set Model Directory to that folder:

modelscope download --model FunAudioLLM/Fun-ASR-Nano-2512 --local_dir ./Fun-ASR-Nano-2512

Advanced STT Settings

• Frame Interval (ms): Determines the real-time responsiveness of STT. A setting of 20ms (near real-time) is recommended. A smaller value means faster text display but higher CPU usage.
• Language: Force a specific language for recognition (e.g., English, Chinese, Japanese). Selecting Auto is recommended to let the AI decide.
• TextNorm: Text normalization processing. Keeping it at auto (default) is recommended.

native-stt Offline File Transcription

The upgraded native-stt path adds an offline transcription workflow alongside real-time microphone STT. It is intended for longer recordings, archived audio, meeting replays, and local file-based transcription.

Best-fit Scenarios and Behavior

• Audio files only: this path is for uploaded/local audio files. For live microphone STT, use ONNX or FunASR backend in real-time mode.
• Native backends: switch between Whisper and SenseVoice based on the model you want to run.
• Safer long-audio handling: long files are processed in chunks and appended incrementally to the Raw Result area.
• Interruptible: you can stop an in-progress transcription with the stop button.
• CUDA preferred when available: if the machine has a usable CUDA environment, native-stt will try GPU acceleration first.

How to Configure native-stt

1. In the STT panel, switch the backend to Whisper or SenseVoice.
2. Select the corresponding model file or model path.
3. Select the audio file you want to transcribe.
4. Click Transcribe File; partial and final results will be appended to the Raw Result area.

Model and Format Notes

• Whisper: commonly used with .bin, .ggml, or .gguf model files.
• SenseVoice: commonly used with .bin or .gguf model files.
• Language: in most cases, auto is recommended; manually locking language may improve stability for some files.