# Text to Speech

> Turn up to 1,500 characters of text into speech with a library or cloned voice. Engines, modes, emotion tags, pauses, voice settings, Arabic text processing and how it is charged.

## What it does

Text to Speech turns a short text into an audio clip: type or paste the text, pick a voice, generate, download. It is made for clips of up to 1,500 characters. For books, long articles or anything longer, use [Audio Studio](audio-studio.md).

## Use cases

Voiceovers for videos and ads, announcements, social clips, e-learning snippets, product audio, and spoken replies in apps through the API.

## Input and output

| | |
| --- | --- |
| Input | Text, up to 1,500 characters per clip, in more than 130 languages |
| Voices | The voice library, your cloned voices, and your team's voices |
| Output in the web app | MP3 or WAV |
| Output through the API | MP3, WAV, PCM or Opus at several sample rates; see [Generate speech](../api/tts-generate.md#output-formats) |

## How to use it

1. Open **Text to Speech**.
2. Click **Select a voice** and choose one. Press play on a voice to hear it first.
3. Type or paste the text.
4. Check the estimated credits next to **Generate**, then click **Generate**.
5. Play the result and download it. Earlier clips are in **History**.

## Engines

A switch at the top of the page selects the engine. Each voice belongs to one engine.

- **Standard**: studio voices, your cloned voices and full voice settings, with three modes.
- **Advanced** (alpha): its own set of voices, directed in plain words through the **Context** box. It is enabled per deployment and may not be available. Its credits carry a multiplier; see [Pricing](../pricing/index.md).

If the Advanced engine stops before the end of the text, Moknah notices and generates the clip again on its own. If it still cannot finish, the credits are refunded.

### Standard engine modes

| Mode | API value | Best for |
| --- | --- | --- |
| Emotion and dialect (default) | `emotion` | Fast, expressive speech with emotions and dialects. |
| Quality emotion and dialect | `quality` | Richer, more natural emotions and dialects. Takes a little longer. |
| Studio Quality | `studio` | Clean, consistent narration with full voice controls. No emotion tags. |

`quality` and `studio` cost more credits than `emotion`. The current multipliers:

| Voice engine and mode | Credit multiplier |
| --- | --- |
| Standard engine, `emotion` mode (default) | x1 |
| Standard engine, `quality` mode | x2 |
| Standard engine, `studio` mode | x2 |
| Advanced engine voices | x1.5 |

### Advanced engine settings

| Setting | What it does |
| --- | --- |
| Context | Describe how the text should be performed: the mood, the pace, who is speaking and to whom. The strongest control. |
| Temperature | Lower makes the delivery steadier and more predictable, higher makes it more varied and expressive. |

## Emotions, sounds and pauses

In the two emotion modes (`emotion` and `quality`) the delivery can be directed with tags in square brackets. Tags are written in English whatever the language of the text.

| Type | Examples |
| --- | --- |
| Emotions | `[happy]` `[excited]` `[sad]` `[angry]` `[afraid]` `[calm]` `[curious]` |
| Delivery | `[whispering]` `[shouting]` `[serious]` `[sarcastic]` `[questioning]` `[crying]` |
| Sounds | `[laughing]` `[chuckles]` `[sighs]` `[gasps]` `[coughs]` `[yawns]` |
| Effects | `[applause]` `[clapping]` `[gunshot]` `[explosion]` |

Pauses: in the web app click **Add pause** and choose a length. In text sent through the API or MCP, write `[break_X]` with X in seconds from 0.1 to 10, for example `[break_1.5]`. At most 20 pauses and 30 seconds of pause per text.

Limitations to tell users about:

- Tags are still under testing and do not work every time. Listen to a tagged line before relying on it.
- Tags do nothing in Studio Quality mode.
- Very short texts can sound uneven. Give the voice at least a sentence or two.

## Voice settings

| Setting | What it does |
| --- | --- |
| Temperature | Higher gives the voice more freedom and expression, lower makes it steadier. Around 25 to 50 is a good start. |
| Similarity | How closely the result follows the original voice. Lower it if a cloned voice came from noisy samples. |
| Speed | Studio Quality only, from 0.7 to 1.2. Extreme values can lower the quality. |
| Expressiveness | Studio Quality only. Adds emotion to the delivery. Keep it low for long narration. |
| Language | Leave it on Auto unless the voice reads the text in the wrong language. |

## Text processing for Arabic

| Option | Use it for |
| --- | --- |
| Basic | Every language, and Arabic dialects. |
| Arabic Tashkeel | Standard Arabic. Adds diacritics so every word is pronounced correctly. |
| AI Poet | Poetry and literary Arabic. |

Arabic Tashkeel and AI Poet are offered only when the content language is Arabic, and they have their own price per character. See [Languages](../platform/languages.md) and [Pricing](../pricing/index.md).

## Mastering

After generating, open **Mastering** in the player. Start from a preset such as Podcast, Warm or Cinema, adjust bass, treble, reverb and loudness, then download the result.

## Writing tips

Numbers and symbols are read more reliably when written out.

| Instead of | Write |
| --- | --- |
| `Dr.` | Doctor |
| `100%` | one hundred percent |
| `moknah.io/docs` | moknah dot io slash docs |

## Credits

Credits depend on the number of characters, the text processing option, the voice (some voices carry a fee) and the mode or engine. The estimate is shown before generating. Rates are in [Pricing](../pricing/index.md).

## Content check

Text sent for conversion is checked automatically. Text that may break the content policy is reviewed by the Moknah team, and text that clearly breaks it is refused.

## Using it from code or an assistant

- REST: [Generate speech](../api/tts-generate.md)
- WebSocket: [Stream speech](../api/tts-stream.md)
- MCP: the `text_to_speech` tool, available on every plan. See [MCP tools](../mcp/tools.md).

## Common problems

| Problem | What to try |
| --- | --- |
| The audio sounds flat | Raise Temperature a little, use an emotion mode with a tag such as `[calm]`, or pick a voice that fits the content. |
| The result is unstable | Lower Temperature and Expressiveness, and give the voice a full sentence or two. |
| A word is pronounced wrong | For Standard Arabic choose Arabic Tashkeel, add diacritics to the unclear word, write numbers and abbreviations as words, or try another voice. |
| The dialect does not come through | Choose a voice that speaks that dialect, use Basic text processing, and use an emotion mode. |
| The same text sounds a little different each time | That is expected: the voice performs the text fresh each time. Lower Temperature for more consistency. |
