- From: James Salsman <j.salsman@bovik.org>
- Date: Fri, 18 Jan 2002 14:13:03 -0800 (PST)
- To: www-voice@w3.org
- Cc: sanjejpi@cse.mrt.ac.lk
Dave Raggett wrote: > On Fri, 18 Jan 2002, Iroshan Sanjeewa wrote: > >> I'm doing a project which takes a voice from a telephone and >> then convert it to text and store on a database in linux >> flatform. If you know how to convert telephone voice to text >> please tell me. Thanks. >> >> shan > > Telephone voice quality isn't good enough for dictation with > the current generation of speech recognizers.... That is not entirely true. Even with very high quality audio, transcription from automatic speech recognition must be carefully proofread, and there is no efficient audio-only interface for doing so that has yet been devised. For example, Speech Machines Corporation: http://www.cybertranscriber.com/ has been going over-the phone transcription services, based on and ASR first-pass with a human proofreader pool, for years. The best ASR dictation solution for Linux is presently ViaVoice -- http://www-3.ibm.com/software/speech/linux/dictation.html -- from IBM, though it is closed-source and has a limited API (for example, it can't return the endpoints of recognized words and the endpoints of their phonemes -- a deficiency shared by VoiceXML 2.0.) Open source speech recognition systems which are more capable, but require speaker-dependent language model training and adaptation for good dictation results, are Entropic HTK and CMU Sphinx: http://htk.eng.cam.ac.uk/ http://www.speech.cs.cmu.edu/sphinx/ Best wishes, James
Received on Friday, 18 January 2002 17:13:05 UTC