github / github/app

Strip filler words from voice dictation before prompt submission

未关闭
#942 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看
Requests and ideas
主要语言
没有语言数据
星标
2.1k
派生
153
PR 合并指标
30 天内没有已合并 PR

描述

🤖 *AI response on behalf of Casey*

### Feature summary

Add an optional post-processing step for GitHub Copilot app voice dictation that removes filler/dead words such as "um", "uh", "ah", "you know", and repeated disfluencies before the prompt is submitted to the model.

### What problem are you trying to solve?

Voice dictation often captures natural speech artifacts that are useful while thinking aloud but not useful in the final prompt. These filler words can make prompts noisier, reduce clarity, and consume extra tokens over time.

Example rough dictation:

> Um, can you uh update the tests, you know, for the new parser behavior?

Cleaner prompt:

> Can you update the tests for the new parser behavior?

### Proposed solution

Before sending dictated text to the model, run lightweight sanitization that can:

- Remove common filler words and repeated verbal pauses.
- Collapse obvious repeated fragments caused by speech-to-text corrections.
- Preserve intentional technical terms and quoted/code text.
- Be configurable or disableable if users prefer exact transcription.

This could be local deterministic cleanup, model-assisted polishing, or part of the existing voice dictation pipeline.

### Workflow impact

This would make voice input more efficient for users who dictate longer prompts, especially in agentic workflows where prompt clarity matters. It should also reduce small amounts of token usage per voice prompt, which could add up across frequent dictation users.

### Related context

Related VS Code backlog request: https://github.com/microsoft/vscode/issues/315508

That request tracks similar speech-input post-processing for VS Code, including removing filler words like "um" and "ah" and reducing invalid/low-value tokens before Copilot processes the input.

贡献指南

打开贡献指南

调研方向

该 issue 没有指出任何 repository 文件、测试或入口点,因此首先在应用中定位语音听写路径和 prompt 提交边界。定义清理范围、可配置性和保留规则,然后添加覆盖,展示填充词移除、重复片段处理,以及提交前技术文本或引用文本保持不变。

由索引模型根据 Issue 内容生成。

评估

技术栈
github
领域
ai, desktop
Issue 类型
功能
难度
5/5
预计耗时
一周以上
活跃度
冷清
描述清晰度
需要澄清
新手友好度
25/100

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。