Thai Speech Dataset for Manufacturing & Shipbuilding Work Instructions

A single-turn Thai speech dataset of work instructions used on shipbuilding and manufacturing sites. Scripts are the instructions actually issued on the floor — equipment inspection, safety briefings, tool handover and storage, measurement recording, progress reporting, consumable distribution — phrased as a supervisor speaking to a worker, and recorded by Thai speakers so that models can recognize commands as they are spoken by the foreign-national workforce common to these sites.

Thai has no letter case, so domain terms are not visually marked; technical measures appear with their standard abbreviation instead, as in the mechanical impact rating (IK). Because on-site speech is dense with equipment vocabulary and delivered as short imperatives, the dataset keeps that phrasing rather than normalizing it into written prose.

Applicable Areas

• Domain-adapted ASR for industrial and shipyard environments
• Multilingual voice interfaces for on-site work instruction
• Speech-to-Speech instruction relay across language barriers
• Safety-critical command recognition and keyword spotting

Collection

Work instruction text is gathered from shipbuilding and manufacturing sites and normalized into single-utterance scripts, with equipment names, tool names, and measurement terms preserved rather than simplified. Scripts are balanced across instruction types — inspection, safety, preparation and placement, recording and reporting, cleanup and return — so that the set is not dominated by any one command pattern. Each script is then recorded by a Thai speaker, and the speaker's gender is stored on the record.

Validation

Every recording is transcribed and checked against the source script for omissions, substitutions, and mispronounced equipment terms, since a single misheard tool name changes the instruction. Reviewers additionally confirm that recording level, clipping, and background noise fall within spec, because deployment is in a noise-heavy environment. Utterances failing transcript match or audio spec are re-recorded rather than corrected in post-processing.