Can AI Voice Input Replace Typing? A Hands‑On Review of Daiti’s Offline Speech‑to‑Text Tool

Background

Frequent typing errors and the need to delete and re‑type reduced productivity when writing AI‑related articles. Built‑in speech recognition performed poorly on mixed Chinese‑English input, and foreign AI voice‑typing solutions were expensive and cumbersome. A Chinese product named “代体” (Daiti) was discovered, promising four‑times faster input, offline operation, and free usage.

Evaluation

Positive aspects

Runs a local speech‑recognition model, providing fast transcription and keeping data private.

Supports a custom vocabulary, improving accuracy for mixed Chinese‑English content.

Activation is a single shortcut: press, speak, release.

AI‑driven text correction fixes homophones, typos, and removes filler words (e.g., “搜狐” corrected to “搜狗”).

Correction parameters—prompt, temperature, max tokens, model name—are user‑configurable.

Recording and transcription history can be reviewed.

Areas for improvement

Custom dictionary must be built manually; large lists may become ineffective.

AI correction currently integrates only Volcano Engine; other providers are unsupported.

No import/export or keyword search for history records.

While speech recognition can stay local, the correction service sends transcriptions to the provider, raising privacy and cost concerns.

Future versions could offer fully offline correction models.

Installation and usage

Download the appropriate version from the official site:

https://daiti.ai/

Installation is straightforward: hold the configured shortcut key, speak, and release.

Populate the custom dictionary with domain‑specific terms to improve recognition. Example prompt for generating a professional term list:

我的职业是 xxx，我想让你帮我生成语音转文字的专业词典，请你帮我生成出 xxx 个行业的常用词语，主要包括：xxxx 这几个方面。
多个词语之间换行，请直接输出不要加序号之类的

When AI correction is enabled, the following system prompt guides the model:

你是一个语音转写文本纠正助手。

你的任务：
- 修正语音识别文本中的识别错误、同音字错误、错别字和标点问题；
- 保持原意，可适当删减不必要的语气词（如“嗯”“啊”）和重复词语；
- 当识别结果中出现与用户词典中词汇发音相似、拼写接近或语义相关的词时，必须将其替换为词典中的标准形式；
- 若识别结果中已正确包含词典词汇，请保持原样；
- 不要更改词典中词汇的拼写、大小写或符号；
- 英文单词和中文之间需要加一个空格；
- 我是做 AI 应用工程师，需要检查专业词汇，避免将“AI”识别成“AR”，将“观猹网”识别为“观察网”等；
- 当出现类似“前面说错了请删除”或“这里说错了，请删除”的指令时，自动删除该句及其前面的内容。

输出：调用一次名为 return_correction 的函数，参数：
status: "ok" 或 "filtered"
text: 纠正后的文本或原文
reason: 可选（若触发内容安全限制，说明原因）

Example input:

“AI 非常有价值[说错了，请删除] AI 能够让我们超越自己的智力和能力极限”。

Corrected output removes the bracketed segment and fixes “AI” misrecognition.

To enable AI correction, obtain an API key from Volcano Engine’s console and configure it in the application. Volcano Engine provides a free quota of 500 000 tokens; beyond that the pricing is 0.4 CNY per 100 k input tokens and 1 CNY per 100 k output tokens.

Performance test: using Sogou input the author’s peak speed was 145 characters/min, while Daiti reached up to 196 characters/min, demonstrating a measurable productivity gain.

Conclusion

Daiti offers a free, offline‑capable voice‑typing solution with customizable dictionaries and AI‑assisted text correction. The workflow—press shortcut, speak, release, optionally enable correction—delivers faster transcription and reduces manual editing. Privacy considerations remain for the correction step, which currently relies on a cloud service, but the overall cost is modest given the free token quota.