# Talk, then edit: dictation and voice mode

Speak the background and goals, type the exact names and numbers, then check the transcript: what dictation does well, where it invents or drops words, and a prompt that turns a ramble into a brief.

Guides · v1.1 · draft · https://hayzlabs.com/guides/talk-then-edit/

In short:
- Talk to give background, goals and a first draft; type exact names, numbers, codes and the final edit.
- Paste the raw transcript into a prompt that turns it into a brief and lists words that look like transcription errors.
- Before you use dictated text, check every number, every name, every not, no and never, and any sentence you do not remember saying.

## Short answer

Speaking is faster than typing on a phone, and it makes a long, specific brief cheap to give. It is not accurate enough to trust without a look. We suggest talking to give background, goals and a first draft, and typing the exact names, numbers, code and the final edit. Then run a short check on the transcript before you use it. The split, the prompt (the text we give the assistant) and the check are our suggestion and untested.

**Dictation** is you speaking while the app types. **Voice mode** is a spoken back-and-forth with the assistant. This page is about dictation. To transcribe a recording of other people, such as an interview or a call, use [Transcribe a recording, then check it](https://hayzlabs.com/workflows/transcribe-then-check/) instead.

## What the evidence says

**Speed.** In a 2016 Stanford lab study (n = 48), speaking was about three times as fast as typing on a phone keyboard in English (2.93x: 153 against 52 words a minute). It used a 2016 speech recogniser and short phrases ([Ruan et al.](https://www.arxiv.org/pdf/1608.07323), lab study). Do not expect the same on a laptop with a good keyboard. As arithmetic and not as your speeds: a 300-word background takes about 2 minutes to say at 153 words a minute and about 6 minutes to type at 52.

**Invented words.** In one study, transcriptions by Whisper (an OpenAI speech-to-text tool) sometimes contained entire invented phrases ([Koenecke et al., FAccT 2024](https://arxiv.org/abs/2402.08021); more in [Transcribe a recording, then check it before you quote it](https://hayzlabs.com/workflows/transcribe-then-check/)). That is one tool in one study, but it means a transcript can look fluent and still say something you did not.

We found no controlled study showing that spoken prompts give better AI output than typed ones. The benefit is that a fuller brief becomes cheaper to give.

## Talk or type

| Talk | Type |
|---|---|
| Background and goals | Exact names, numbers, codes and addresses |
| A first draft of a note or an email | The final edit |
| Describing what you can see on screen | Anything private that others can overhear |
| When you are away from the desk, walking or cooking | Short commands such as "shorter" or "redo point 3" |

Spoken drafts run long and loose, which is fine for a first pass. Ask for a cut before you read.

## The routine: ramble, structure, edit

1. **Ramble.** Open a notes app or the chat box and tap the microphone on your phone's keyboard (in Gboard on Android, tap Microphone at the top of the keyboard and wait for "Speak now"). Say the goal, the background, the constraints and what you are unsure of. Do not tidy as you go. Two or three minutes is plenty.
2. **Structure.** Paste the transcript into the prompt below and let the assistant turn it into a brief. This is the step that makes a ramble usable. For what a good prompt holds, see [Prompting in one page](https://hayzlabs.com/start/prompting-in-one-page/).
3. **Edit on screen.** Type the exact names, numbers and the final wording. Run the check below.

```text title="Ramble to brief · v1.0"
Below is a rough voice transcript. It has filler words and may have
transcription errors.
1. Turn it into a brief with four headings: Goal, Background, Constraints,
   Open questions.
2. Keep my wording for anything that sounds like a decision or a number.
3. List any words that look like transcription errors (names, numbers,
   jargon) so that I can check them.
Do not add facts I did not say.
Transcript:
<TRANSCRIPT>
```

Replace the following:
- `<TRANSCRIPT>`: the raw dictated text, pasted as it came out.

> [!CHECK]
> Before you use anything dictated, scan it for four things: every number, every name, the words "not", "no" and "never", and any sentence you do not remember saying. Transcription errors look fluent, and a dropped "not" reverses the meaning. Read these four things against what you meant, and not against the transcript.

**Example (illustrative).** Ines, a freelance translator, dictates a new client's style notes on her walk to the studio. At her desk she runs the prompt, checks the client name and the three numbers, and finds that "not formal" had come out as "formal". She types the corrected line.

## How to tell it worked

- The brief has all four headings, and nothing in it is a fact you did not say.
- Every number and name matches what you meant, and every "not" is where you put it.
- It has failed if dictated text went into a message or a file before the four-part check.

## Where your voice goes

What happens to the audio, and whether it is used to train the AI, depends on the app and your settings, and it changes. Do not dictate anything you would not paste: [What not to paste](https://hayzlabs.com/start/what-not-to-paste/). [Privacy settings pass](https://hayzlabs.com/setup/privacy-settings-pass/) covers the settings. In a shared room, type.

> [!LIMIT]
> The speed result is from 2016, on a phone, with short test phrases and 48 people. Recognisers have changed, and on a laptop the gap may be small or reversed. The invented-phrase result is for one tool in one study. Speech differences can raise error rates, and that is not the speaker's fault. Nothing here shows that spoken prompts produce better answers. For anything a client, a court or a doctor will rely on, check the transcript word by word.

## Sources

1. Ruan, Wobbrock, Liou, Ng & Landay. Speech and phone keyboard input, 2016. Lab study, 48 people. [Link](https://www.arxiv.org/pdf/1608.07323)
2. Koenecke et al. Invented phrases in Whisper transcriptions, FAccT 2024. Peer-reviewed benchmark. [Link](https://arxiv.org/abs/2402.08021)