Skip to content

Feature Guide English

Jump to bottom
zkwi edited this page Apr 27, 2026 · 4 revisions

Features and Usage Optimization

This page explains what VoxType does, when each feature is useful, and how to tune it for speed, latency, and recognition quality.

简体中文版本：功能特性与使用优化

1. Core Workflow

Trigger recording → Capture microphone audio → Doubao streaming ASR → Optional LLM polishing → Write clipboard → Auto paste → Restore clipboard → Stats and logs

Main workflow guarantees:

  • Empty recognition becomes a failure and does not trigger polishing, paste, or successful statistics.
  • The UI only shows “polishing” when LLM polishing is enabled, text reaches min_chars, and Base URL, API Key, and model are complete.
  • Floating captions show real-time text, elapsed state, and errors only.
  • Usage statistics do not store recognized text.

2. Home

The Home page is for daily use:

  • Check current input state.
  • Start or stop recording.
  • See main and backup trigger states.
  • See recent 24-hour, 7-day, speed, and saved time estimates.
  • After a successful input, temporarily view and copy the latest recognized text in the current window. It is not written into usage stats, logs, or diagnostic reports, and is cleared when the window closes or a new recording starts.

3. Global Triggers

Default trigger: Ctrl + Q.

Trigger Default Recommendation
Main shortcut On Keep it on
Middle mouse Off Enable only after checking browser/editor conflicts
Right Alt Off Enable only after checking IME/system shortcut conflicts

Avoid enabling multiple easy-to-misfire triggers unless you really need them.

4. Microphone Capture

VoxType uses Rust cpal to capture PCM audio.

Defaults. These are internal defaults, so most users do not need to change them in the settings UI:

  • Sample rate: 16000
  • Channels: 1
  • Segment size: 200ms
  • Max recording duration: 300s
  • Local silence auto-stop: 10s

Tips:

  • Select a fixed input device if you have multiple microphones.
  • Low volume, long distance, noisy rooms, and disabled Windows microphone permission can cause empty recognition.
  • If a recording contains no useful speech, VoxType uses the local silence fallback to stop recording instead of waiting for the maximum duration.
  • System-volume mute while recording is off by default; enable it only if echo affects recognition.

5. Doubao Streaming ASR

VoxType uses Doubao bigmodel_async WebSocket by default.

It supports real-time partial text and final recognition. Low-level ASR request fields remain supported in config.toml, but ordinary users do not need to edit them.

Quality and latency factors:

Factor Recommendation
Audio segment Keep the internal default 200ms
Server endpointing end_window_size defaults to 800ms; existing manual config is preserved
Local silence fallback Continuous silence stops recording after 10 seconds by default
Final result timeout Default 15s; adjust only for network/service issues
Hotwords Important for proper nouns and product names
Recent context Useful for continuous writing, but disabled by default for privacy

6. Auto Paste and Clipboard Restore

After recognition, VoxType:

  1. Writes the final text to the clipboard.
  2. Sends Ctrl+V or Shift+Insert.
  3. Tries to restore the previous clipboard.

Default UI exposes:

  • Auto paste: recommended.
  • Clipboard only: useful when you do not want VoxType to send paste keys.

Advanced settings expose:

  • Full paste mode: ctrl_v / shift_insert / clipboard_only
  • Restore clipboard after paste
  • Clipboard restore delay

If paste fails but text was copied, press Ctrl + V manually.

7. LLM Polishing

LLM polishing is useful for:

  • Cleaning dictated long sentences.
  • Removing filler words and repetition.
  • Structuring long text into paragraphs or bullet-like lines.
  • Preserving proper nouns with hotwords and scene notes.

Defaults:

  • ASR only unless LLM polishing is enabled.
  • Short text below min_chars is not polished.
  • Thinking is disabled to reduce latency.

Tips:

  • Short messages often do not need polishing.
  • Documentation, meeting notes, and requirement drafts benefit more from polishing.
  • If polishing is slow, disable thinking, raise min_chars, or choose a faster model.

8. Hotwords, Scene Notes, and Prompts

Use hotwords for terms that ASR often misrecognizes:

VoxType
Tauri
Doubao ASR
Project code names
Product names

Use scene notes for long-term style and context:

I often dictate product requirements, code review notes, and project plans.
Keep output concise and do not expand.
Preserve English technical terms and project names.

Prompt editing:

  • User Prompt template is available from the default page.
  • System Prompt and minimum polishing length are advanced settings.
  • Final prompt preview is available.

9. Automatic Hotword Candidates

Automatic hotword candidates extract possible terms from local voice input history.

Privacy behavior:

  • Off by default.
  • Stores only VoxType final voice input text.
  • Does not record keyboard input.
  • Does not read clipboard history.
  • Calls the configured LLM only when you manually generate candidates.
  • Candidates must be confirmed before joining hotwords.

Enable it when you frequently dictate recurring business terms, product names, or people names. Keep it off when your dictated content is highly sensitive.

10. Floating Captions

Captions are designed for recording feedback:

  • Real-time text.
  • Elapsed or processing state.
  • Errors.

Captions do not show paste-state noise, internal paths, or debug stacks.

Use presets first. Fine-tune width, height, colors, and bottom margin only if captions block your content.

11. Statistics

Stats record:

  • Session count
  • Voice duration
  • Character count
  • Average speed
  • Saved time estimate

Stats do not store recognized text.

12. Tray, Startup, and Updates

Tray:

  • Closing the main window hides it to the tray by default.
  • Tray menu can open the main window, open config, open logs, and exit.

Startup:

  • Can be enabled from Options.

Updates:

  • Manual update check is in advanced settings.
  • Startup auto-check is enabled by default.
  • When a new version is available, notices and the update panel provide an "Update now" action.
  • Updates download the NSIS installer from GitHub Releases, exit the current app to release files, and try to open the new version after installation.

13. Latency Optimization

Goal Tune first
Faster real-time captions Keep the internal 200ms audio segments and ensure network stability
Faster final output Disable LLM polishing or increase min_chars
Faster polishing Disable thinking and choose a faster model
More reliable paste Keep clipboard restore enabled and increase restore delay if needed
Fewer accidental triggers Keep right Alt and middle mouse disabled
Better proper nouns Maintain hotwords and scene notes

简体中文

English

Clone this wiki locally