CapsWriter-Offline (v2.5)

Hold CapsLock, speak, release, and the text appears.

CapsWriter-Offline is a fully offline speech input tool built primarily for Windows.

🚀 What’s New

v2.5-alpha

• Initial support for Qwen3-ASR-1.7B
  • Works for both microphone dictation and file transcription in this fork
  • When the model does not return real timestamps, the app falls back to the existing approximate timeline path for subtitle and JSON generation
  • Decoder-side Vulkan acceleration is enabled by default and typically needs about 1.6 GB of VRAM
  • If your GPU drops memory clocks while idle, cold-start latency can rise to around 300 ms
  • Locking memory clocks with nvidia-smi -lmc 9000 can reduce short-clip latency to around 100 ms on hardware such as RTX 5050

v2.4

• Improved Fun-ASR-Nano-GGUF support with DirectML encoder acceleration and better FP16 defaults
• Server-side Fun-ASR-Nano now uses its own hot-server.txt hotword context file
• Spoken punctuation like “comma”, “period”, and “new line” can be converted automatically
• Added decoder temperature handling to avoid edge-case repetition loops
• Improved server-side alphabet spelling merge behavior

v2.3

• Added Fun-ASR-Nano-GGUF support
• Refactored large-file transcription to async streaming
• Improved Chinese/English spacing cleanup
• Improved server cleanup after abnormal disconnects

✨ Core Features

• Speech input: hold CapsLock or mouse side button X2, speak, release, and insert text immediately
• File transcription: drag audio/video files onto the client and generate .srt, .txt, and .json
• ITN formatting: convert spoken number patterns into clean written forms
• Server hotword context: store domain-specific terms in hot-server.txt to help Fun-ASR-Nano recognize context
• Hotword replacement: use hot.txt for phoneme-based fuzzy matching and forced replacement
• Rule replacement: use hot-rule.txt for regex-based or direct text replacement
• Rectify history: keep correction history in hot-rectify.txt to help LLM polishing
• LLM roles: route text to roles such as assistant or translate when the recognized text starts with that role name
• Tray menu: manage hotwords, copy results, or clear LLM memory from the tray icon
• Client/server split: run the model on one machine and the lightweight client on another if needed
• Diary archive: save recognized sentences by date
• Audio archive: save recorded audio locally for privacy and traceability

The project is designed around four ideas: offline, fast, accurate, and highly configurable. The goal is a smooth voice-input workflow that still works without cloud access, installation-heavy deployment, or network connectivity.

LLM roles can use local models through Ollama or remote APIs through providers such as OpenAI-compatible services.

💻 Platform Support

The project is mainly targeted at Windows 10/11 (64-bit).

• Linux: not officially tested or packaged
• macOS: currently unsupported because low-level keyboard-hook support is limited and system permissions are restrictive

🎬 Quick Start

1. Install the VC++ runtime.
2. Download the app from Latest Release.
3. Download the model package from Models Release and extract it into the matching subfolder under models/.
4. Launch start_server.exe.
5. Launch start_client.exe.
6. Hold CapsLock or mouse side button X2 and start speaking.

🎤 Model Options

Select the speech model in config.toml through server.model_type:

• qwen3_asr: built-in punctuation, acceptable CPU speed, very fast on discrete GPUs, excellent accuracy
• fun_asr_nano: built-in punctuation, fast on CPU, very fast on discrete GPUs, top-tier accuracy
• sensevoice: built-in punctuation, extremely fast on CPU, strong multilingual support
• paraformer: external punctuation model, extremely fast on CPU, high accuracy

⚙️ Configuration

All runtime settings live in the root config.toml.

• Edit [[client.shortcuts]] to change keyboard or mouse triggers
• Set hold_mode = false for press-once / press-again recording
• Toggle llm_enabled to enable or disable LLM post-processing
• Change server.model_type to switch ASR backends
• Tune model-specific acceleration flags under [models.*]

🛠️ FAQ

Q: Why does nothing happen when I press the key?
A: Make sure the start_client.exe console process is still running. If you want to type into an elevated application, run the client with administrator privileges too.

Q: Why is there no recognition output?
A: Check the recorded audio inside the dated year/month/assets folder. Make sure the microphone is actually recording and that Windows microphone permissions are enabled.

Q: Can I use GPU acceleration?
A: Fun-ASR-Nano and Qwen3-ASR can use GPU acceleration. If your integrated GPU performs worse than CPU, disable the model-specific dml_enable or vulkan_enable flags in config.toml.

Q: File transcription is too slow on low-end hardware. What can I do?
A: Try these in order:

1. Use sensevoice or paraformer if you need the fastest CPU path.
2. Disable dml_enable or vulkan_enable for Qwen3-ASR / Fun-ASR-Nano if GPU acceleration hurts more than it helps.
3. Lower model thread counts where supported.
4. Lock GPU memory clocks with nvidia-smi -lmc 9000 if you are optimizing for short clips on NVIDIA hardware.

Q: Fun-ASR-Nano quality is unstable on my integrated GPU. Why?
A: Some integrated GPUs have poor behavior with Vulkan FP16 accumulation in llama.cpp. If you see degraded output, disable vulkan_enable for that model and run the decoder on CPU.

Q: How do hotwords work?
A: hot-server.txt is used for server-side Fun-ASR-Nano context enhancement. hot.txt and hot-rule.txt are client-side replacement sources. hot-rectify.txt stores correction history used by the LLM pipeline.

Q: How do I use LLM roles?
A: Start your spoken command with the role name. For example, if you have a role named translate, saying translate, the weather is great today sends the recognized text through the translation role instead of direct output.

Q: How do I choose the LLM model behind a role?
A: Role defaults and overrides are defined through config.toml plus the LLM/ role entry modules. You can point roles to local Ollama models or remote API providers.

Q: Can an LLM role read selected text on screen?
A: Yes. If a role enables enable_read_selection, the app can capture the current selection with Ctrl+C and pass it into the LLM context before processing your voice command.

Q: How do I hide the console window?
A: Use the tray menu to hide it.

Q: How do I start it automatically on boot?
A: Run Win+R, open shell:startup, and place shortcuts for the client and server there.

❤️ Credits

This project builds on several excellent open-source projects:

• Sherpa-ONNX
• FunASR

Special thanks to modern AI coding assistants and to the users who supported the project.