IBM Corp, which has been playing with speech recognition for years and years and years, now has the technology down to a state where it will run on an 80486-based personal computer rather than the mainframes on which it started, and has launched the IBM Personal Dictation System, which costs $500 for the board and another $500 for the software. It ships December 28 and follows the system for the RS/6000 launched last year. The system is claimed to offer the most accurate large vocabulary speech recognition capability yet, with a 32,000-word vocabulary and the ability to take dictation at up to 70 words per minute. There are also language modules for different professions such as journalists, emergency medical practitioners and radiologists, which cost between $400 and $500, and a legal language model is on the way. It needs OS/2 2.1 and works with WordPerfect for OS/2 5.2, WordPerfect for Windows 5.2 and Microsoft Corp’s Word 1.1 for OS/2. Users can also create their own spoken macros for regularly-performed functions. Where it doesn’t recognise a word, it stores it accoustically, the user then types it in after the end of the dictation session, and the system adds it to the vocabulary. The user has to train it by reading a Mark Twain short story for half an hour. It is claimed to distinguish between homophones such as to, two and too, recognize the start of a sentence and capitalise the first letter. It’s US English only at present, British English, French, German, Italian and Spanish are promised by June. It needs 8Mb of higher memory and 32Mb of free disk space – and another 30Mb of disk is needed for the training session.