On this page

Text-to-Speech (TTS)

Speaks any text the pipeline produced back onto the asset.

The TTS node generates narration audio from text. Unlike the analysis nodes, which describe a property of the media itself, TTS has a single text input port that accepts any text type — connect an LLM summary, a caption, or a translated transcript to it — and synthesizes speech, attaching the audio to the asset.

German is synthesized with Orpheus-3B / Kartoffel and English with Kokoro. Both run behind a small HTTP sidecar, so the node stays a pure client and needs no model runtime of its own.

Kind

tts

Applies to

Any asset (the audio is generated from upstream text, not from the media)

Input ports

text — accepts any text type, so an LLM answer, a caption or a transcript can be wired straight into it

Output ports

audio (the generated WAV, ready for S3 Sink), flag (String)

Requirements

A running TTS sidecar (/v1/tts). German uses Orpheus/Kartoffel (GPU recommended); English uses Kokoro (CPU).

Persists to

asset_node_result ledger only — the generated WAV stays in the worker’s local tts_bin cache (as with thumbnails)

Configuration

The Text to Speech settings panel: Sidecar Host
Figure 1. The node’s settings in the pipeline editor

Set these in the panel above, or in the node’s options block in a pipeline definition:

Option Meaning

ttsHost / ttsPort

Address of the TTS sidecar (default localhost:9100)

language

de → Orpheus/Kartoffel, en → Kokoro (default de)

voice

Voice id for the selected engine (default Jakob; English e.g. af_heart)

Nothing configures where the words come from — connect a port to text and the node speaks whatever arrives on it.

Seeing it run

Turn on Debug Mode and every node keeps what it produced, on the card itself. Below is a real run at the shipped language: de / voice: Jakob, speaking the German transcript that Translate and Sentiment also work on.

The TTS node after a run

audio carries the path of the WAV in the worker’s local cache and flag records how the node finished. The audio itself is not uploaded anywhere yet — the bytes live beside the worker, exactly as Thumbnail leaves its contact sheet.

Use Cases

  • Audio descriptions — speak an LLM-generated description or caption for accessibility.

  • Dubbing / narration — voice a translated transcript back onto the asset.

  • Bilingual output — German via Orpheus/Kartoffel, English via Kokoro, chosen per pipeline.

The sidecar is a small FastAPI service; see the server/ directory next to the node for how to run it (including the production path where Orpheus runs on vLLM / llama.cpp).

Looking for something else?

All documentation Feature overview