📄 TELÓ: AI-Driven Automatic Subtitling for the Promotion of the Performing Arts

Summary

Theatre performed in Catalan rarely tours abroad, since few foreign spectators understand the language and traditional surtitling needs a translation prepared in advance plus an operator cueing lines live. The TELÓ project (UOC, funded by the Generalitat de Catalunya) proposes an open-source framework that automates this for Catalan, Spanish, English and French in any direction. A cascaded pipeline links four modules: a Transcriber (Whisper, a Catalan-tuned Whisper, or commercial ASR APIs); a Synchronizer that aligns live transcripts with the script, times subtitles and detects improvisation; a Translator based on MTUOC, using NMT or LLMs for scripts and ad-libs; and an Emissor serving per-language streams to phones, tablets, smart glasses or projectors. A local Wi-Fi intranet keeps latency under one second. As a project description, the paper reports no quantitative evaluation: the architecture is built and preliminary tests done, with professional pilots due by August 2026. The work supports internationalisation of minority-language drama and captioning for DHH audiences.

Key points

• Motivation from sector data: Catalonia has 787 venues and 13,647 performances per year. Non-verbal shows make up 8.33% of local performances but 60.59% of international tours. Spoken-Catalan shows fall from 60% locally to 12.38% abroad.
• Project details: reference PLG002/25/000052, running six months (02/02/2026–01/08/2026). It builds on work from the Aina Project and the Grial Research Group.
• First deployment: a French-language production will be subtitled simultaneously into Catalan and Spanish, which justifies the bidirectional design.
• Interface features: device-specific font and contrast settings, and a "Theater Mode" with low-emission colours to limit light pollution. Same-language captions serve DHH spectators.
• Release and plans: code will be published on GitHub under GNU-GPL v.3. Future work covers MT adapted to dramatic registers, better robustness to noise and improvisation, and a possible end-to-end speech-to-text translation model replacing the ASR+NMT cascade.