---
name: voice-reference-clip
description: >
  Make a reference clip for an Exponential custom voice: pick 8-15 seconds of
  calm, clean speech, cut it to a mono WAV, write its exact transcript, and
  enter both in Settings → Voice. Use when a member wants Exponential's
  spoken replies (English or German) in their own voice (or one they have
  consent to use).
---

# A reference clip for your Exponential voice

Exponential speaks replies in English and German, both cloned from ONE short
clip: no training and no upload. The clip stays on your voice device, and
Settings → Voice only stores its path and its transcript. With the Voice file
empty, each language uses its built-in clip.

Speech out runs on the voice device's GPU only (at least 4 GB). Without one,
replies are shown, not spoken, in either language.

Human version: https://exponential.at/docs/libraries/etalk/guides/your-voice/

## Consent first

Use your own voice, or a voice whose owner agreed to this use. Do not clone
anyone without consent.

## 1. Pick the take

- 8 to 15 seconds of ONE speaker, speaking calmly in the language picked in
  Settings → Voice.
- Clean speech: no music, no background noise.

## 2. Cut it to a WAV

With ffmpeg (start time and duration in seconds):

```bash
mkdir -p ~/voice
ffmpeg -ss 12.5 -t 10 -i recording.m4a -ac 1 -ar 24000 -c:a pcm_s16le ~/voice/ref.wav
```

- Mono, 24 kHz, 16-bit PCM (the built-in clips' format).
- Start and end on silence, not mid-word.
- Listen to the result once, all the way through.

## 3. Write the transcript

Type exactly what the clip says, word for word, with normal punctuation. Leave
out filler you can't hear clearly.

## 4. Enter it in Exponential

In Settings → Voice, pick your desktop as the voice device. Then:

- **Language:** English or Deutsch.
- **Voice file:** the WAV's full path on that desktop, e.g.
  `/home/you/voice/ref.wav`.
- **Transcript:** the text from step 3.

The voice device accepts only an absolute path to a local `.wav` of at most
10 MB, with a transcript: no URL, no `..`, on Windows no network share.
Anything else is ignored and the built-in voice speaks. Leave Voice file empty
for the built-in voice.