Die Bach is a robot that tours primary schools in the Rhondda talking to Key Stage 2 children about compound semiconductors and wildlife conservation. He is puppeted by a performer, and he answers questions out loud, in Welsh or English, in a voice that matches the one used in the live shows.

Wild Connect built the puppet. I built everything that lets him hear, think and speak.

In August 2026 the programme was named a finalist for Educational Programme of the Year at the Wales STEM Awards.

The hard part is the room

A school hall is close to the worst listening environment there is. Forty children, a hard floor, no acoustic treatment, and a performer speaking over the top of the very system trying to hear the audience. Most of the work went into the microphone and audio chain rather than into anything visible.

The compute and audio, before it went back inside the shell

The other constraint is latency. A child asks a question and then waits, and a pause that would be unremarkable in a chat interface is fatal in front of an audience, they lose interest, or worse, they decide he is broken. The target was under 500 milliseconds from the end of speech to the start of the reply, which rules out the obvious architecture of waiting for a full transcription, then a full completion, then a full render of speech. Every stage streams into the next.

Speaking as himself

The voice is a clone of Sion Russell Jones, who also performs Die Bach's singing voice. That is the point of it: the puppet and the AI have to be the same character, or the illusion collapses the first time he answers a question in a stranger's voice.

Bilingual throughout. Welsh is not a translation layer bolted on afterwards, a Welsh-medium school gets a Welsh-speaking robot.

Running where there is no wifi

Schools cannot be relied on for a network connection, and a robot that stops working because the guest wifi is captive-portalled is not a robot anyone will book twice. He runs on his own hardware, a Jetson and a mini PC, with an offline mode and a systemd unit that brings him back up on his own if anything falls over mid-performance.

Set up at a school, alongside the monitoring work he talks about

Karaoke

He sings. There is a mode where the lyrics go up on the screen behind him and the hall sings along, which is the part the children remember and the part that gets them onto the topic in the first place.

Every face is mosaicked. The children are somebody else's to consent for, and the singing is the part worth showing, turn the sound on