Expressive TTS Services

Overview

A TTS service performs expression tags through its expression adapter, which declares what the service performs and writes each tag in the provider’s markup. The services below have one for the models or voices listed. Any other TTS service, or model, speaks only the words: tags are left out of its speech. For a walkthrough, see Expressive Conversation.

Supported Services

Service	Models or voices	Sounds	Deliveries	Pauses (ms)	Emphasis	Delivery changes
`AzureTTSService`, `AzureHttpTTSService`	DragonHD and DragonHD Omni voices	Listed	Listed			Anywhere
`CartesiaTTSService`, `CartesiaHttpTTSService`	All	Laughter	Listed	100–5000		Between sentences
`ElevenLabsHttpTTSService`	`eleven_v3`, `eleven_v4` models	Listed or free text	Listed or free text	100–3000	Yes	Anywhere
`ElevenLabsDialogueTTSService`	All	Listed or free text	Listed or free text	100–3000	Yes	Anywhere
`FishAudioTTSService`	S2 models	Listed or free text	Listed or free text	100–1500	Yes	Anywhere
`GeminiTTSService`	Gemini 3.8 TTS models, on the Gemini API	Listed	Free text, `slow` and `fast`	100–2000	Yes	Between sentences
`InworldTTSService`, `InworldHttpTTSService`	`inworld-tts-2`	Listed	Free text, `slow` and `fast`	100–10000	Yes	Anywhere
`InworldTTSService`, `InworldHttpTTSService`	`inworld-tts-2-flash`	Listed		100–10000		Anywhere
`MiniMaxHttpTTSService`	`speech-2.8` models	Listed	Listed	10–99990		Between sentences
`MistralTTSService`	Voxtral 3 models, such as `voxtral-mini-tts-3-0`	Listed	Listed			Anywhere
`SonioxTTSService`	`tts-rt-v2`	Listed	Listed	100–2000	Yes	Anywhere

Listed: the service performs the names it declares, the provider’s own, which the LLM is offered by their universal names. See the provider’s documentation for its list.
Free text: the service also takes a delivery or sound described in a few words, so any name is passed through as written.
Pauses: a pause outside the range is clamped to it. A service without a range performs no pauses.
Delivery changes: where in a reply the delivery can change. Between sentences, a sentence is spoken in the delivery its words start in, and a span’s whole sentence is spoken in the span’s delivery.

The expression adapter is chosen by the service’s model or voice, so a settings change can change what the service performs, and the LLM is taught the new tags.

Avatars

Tavus

TavusTransport sets the replica’s facial emotion to follow the delivery of the bot’s speech, as each word is spoken. The replica shows the emotion of the delivery’s first name other than slow, fast and emphasis, such as happiness for ["warm", "slow"]. The voice’s baseline, and a delivery not in the table, show neutral. At start, a warning names the policy’s deliveries that have no emotion.

Delivery	`TavusEmotion`
`calm`, `curious`, `confident`	`NEUTRAL`
`warm`, `excited`, `grateful`, `happy`	`HAPPINESS`
`sympathetic`, `apologetic`, `sad`	`SADNESS`
`surprised`	`SURPRISE`
`angry`	`ANGER`
`fearful`, `scared`	`FEAR`
`disgusted`	`DISGUST`
`contemptuous`	`CONTEMPT`

from pipecat.transports.tavus.transport import TAVUS_DELIVERY_EMOTIONS, TavusEmotion

TAVUS_DELIVERY_EMOTIONS holds this table, as dict[str, TavusEmotion].

ExpressionCapabilities

What a TTS service performs, as its expression adapter declares it and as TTSMetadataFrame.expression carries it at start.

from pipecat.expressive.expression import (
    DeliveryScope,
    ExpressionCapabilities,
    ExpressionTerm,
)

Field	Type	Description
`sounds`	`list[ExpressionTerm]`	Sounds the service makes, by its own names.
`delivery`	`list[ExpressionTerm]`	Deliveries the service applies to the words after them, by its own names.
`allows_free_text_sounds`	`bool`	Whether the service also takes sounds outside `sounds`, described in a few words. Defaults to `False`.
`allows_free_text_delivery`	`bool`	Whether the service also takes deliveries outside `delivery`. Defaults to `False`.
`pause_range_ms`	`tuple[int, int] \| None`	The shortest and longest pause, in milliseconds, or `None` when the service can’t pause.
`emphasis`	`bool`	Whether the service stresses words marked `^^word^^`. Defaults to `False`.
`delivery_scope`	`DeliveryScope`	Where in a reply the delivery can change: `ANYWHERE` (the default) or `SENTENCE`.
`service_vocabulary`	`dict[ExpressionDelivery \| ExpressionSound, list[str]]`	The service’s names for universal names it calls something else, best first, such as `{ExpressionDelivery.WARM: ["warmly"]}`. A universal name left out is sent as written.

ExpressionTerm is one sound or delivery: its name, by the service’s own name, and a description of when to use it, which the LLM is shown, or "".

delivery_name

delivery_name(name: str) -> str | None

Returns the name the service performs a delivery by, given the name written in its tag, or None when it can’t perform it.

sound_name

sound_name(name: str) -> str | None

Returns the name the service performs a sound by, or None when it can’t perform it.

Adding an Adapter

To make another TTS service expressive, subclass BaseExpressionAdapter and set it as the service’s expression_adapter_class, or override get_expression_adapter() to choose an adapter by model or voice. The service calls the adapter only with what its capabilities say it performs, by the service’s own names.

from pipecat.adapters.tts.base_expression_adapter import BaseExpressionAdapter

Method	Returns
`capabilities()`	The service’s `ExpressionCapabilities`, or `None` for none.
`format_text(text, delivery, previous_delivery)`	The words to send, in their delivery. `previous_delivery` is that of the words before them, or `None` for a reply’s first words.
`format_sound(name)`	The markup for a sound at its place in the text.
`format_pause(duration_ms)`	The markup for a pause, its length within the service’s range.
`format_end(delivery)`	Text to end a sentence’s request with, for markup that would otherwise last past it, or `""`.
`request_settings(delivery)`	Settings to send beside a request’s text, for a service that takes some expression as settings, such as a speed, or `{}`.

Set the markup apart from the words with spaces, so the word timestamps a service reports line up with the text it was sent.

Get Started

Standards

Pipecat Server

Client SDKs

Pipecat Flows

Pipecat Cloud

CLI

Pipecat Context Hub

Expressive TTS Services

Overview

Supported Services

Avatars

Tavus

ExpressionCapabilities

delivery_name

sound_name

Adding an Adapter

​Overview

​Supported Services

​Avatars

​Tavus

​ExpressionCapabilities

​delivery_name

​sound_name

​Adding an Adapter

Overview

Supported Services

Avatars

Tavus

ExpressionCapabilities

delivery_name

sound_name

Adding an Adapter