Run the read-aloud engine

Provide and run the speech engine (VOICEVOX’s Zundamon) that members hear replies read aloud with. The reading dictionary that applies to every member is kept on the same screen.

How it works

What you do	What happens
Provide a VOICEVOX engine the Control Plane can reach	Japanese read-aloud speaks in Zundamon’s voice
Set up Amazon Polly	Polly reads in its place while the engine is stopped or getting ready
Choose the engine’s mode under Read aloud	The engine starts and stops. On demand, it follows the demand by itself
Write readings into the Tenant-wide reading dictionary	Every member’s read-aloud says that spelling with that reading

Whether to use read-aloud, and which voice, each member decides (Use read-aloud). A speech engine cannot be borrowed from another deployment.

Provide the engine

The VOICEVOX engine is not in the package. Provide it for your deployment target.

Target	What to provide
AWS ECS	The speech engine stack (50-tts). It normally runs at zero and is started and stopped from this screen (ECS runbook)
native, WSL	Run VOICEVOX on Windows, or VOICEVOX ENGINE on Linux (native runbook)
Others	Run a VOICEVOX ENGINE the Control Plane can reach over HTTP

• The Control Plane’s AF_VOICEVOX_URL says where the engine is. The default is http://127.0.0.1:50021
• For Polly, set AF_POLLY_REGION and AWS credentials the Control Plane’s process can see
• While the engine is down, only speech fails. When it is back, read-aloud recovers from the next sentence

Choose the engine’s mode

Open Admin from the account menu and choose Read aloud under Deployment. Only a deployment administrator (super_admin) can open it.

Mode	What happens
Disabled	Read-aloud switches to Polly at once. On AWS, the ECS desired count goes to 0 about a minute later
On demand	The engine starts by itself once demand builds up (2,000 characters in 5 minutes), and stops by itself after 30 minutes with nobody listening
Always on	The engine is never stopped

• Starting takes 1–2 minutes. Until it is ready, Polly reads Japanese in its place. Without Polly there is no sound in the meantime
• Every automatic start and stop is written to the audit log
• The Engine: line shows the engine’s state, whether it is ECS-managed or externally managed, and whether Polly is available

For an engine ECS does not manage (a standalone VOICEVOX, say), the screen changes as follows.

Engine	Screen
An externally managed engine is reachable	There is no On demand. The switch only opens and closes the route; the engine itself is never started or stopped
There is no engine	There is nothing to route to, so the switch stays at Disabled. Once an engine is provided it can be used

Build the pronunciation dictionary

Write the readings of product names and in-house terms into the Tenant-wide reading dictionary and press Save.

# Product names
agent-fleet=エージェントフリート
Webshop=ウェブショップ

• One spelling=reading per line. A line starting with # is a comment
• It applies to every member’s read-aloud. When a member’s own dictionary has the same spelling, the member’s wins
• Other members get a saved change from their next Console load

Tenant administrators and members may ask you to add words that keep being misread.

When it goes wrong

Symptom	Cause and fix
The voice is Polly, not Zundamon	The engine is starting or stopped. On demand, Polly reads until demand builds up and the engine starts
There is no sound at all	The engine is stopped and Polly is not set up. Provide one of them
The switch is stuck at Disabled	No engine is reachable from the Control Plane. Run one at AF_VOICEVOX_URL
A dictionary change does not change the reading	Other members get it from their next Console load. A member’s own dictionary with the same spelling wins

Related

• Use read-aloud: the member’s setting
• Run inference engines on your own GPUs
• Read the audit log