-----Original Message-----
From: owner-rsgb_lf_group@blacksheep.org [mailto:owner-rsgb_lf_group@blacksheep.org] On Behalf Of G4gvw@aol.com
Sent: 04 August 2004 13:05
To: rsgb_lf_group@blacksheep.org
Subject: Re: LF: CW and other modes

As a start John, we might have to define our terms as to what is acceptable as "speech transmission"!

Dependant upon the defined term we can commence with a series of logical steps toward such a goal.

At the "starter" end there is perhaps the simple "Voice-to-code, Code-to-voice" protocol of speech recognition to keyer …

Dear LF Group,

In my rather limited knowledge, there seem to be two approaches. One is to encode the speech waveform in an efficient way that allows transmission at a relatively low data rate, whilst still allowing a reasonable facsimile of the speaker’s voice to be reconstructed at the receiving end. An example of this would be the “CELP” encoding used for GSM mobile phones. I think the problem for LF would be that the minimum practical bandwidth would still be fairly large – I recall G4JNT did some work on digital speech for HF in a SSB bandwidth – perhaps he could comment.

The other way would be not to worry about exactly reproducing the speaker’s voice, accepting that whoever the speaker was, they would always sound like Prof. Stephen Hawking at the receiver, but as G4GVW says, transmitting just the meaning of the spoken words. One way to to do this would be for speech recognition software at the transmitter to generate text that could be transmitted at maybe several 10s to a few 100 bits/s, and then fed into a speech synthesizer at the receiver. I think the problem with this is that speech recognition algorithms don’t seem to be very accurate – when used with word processors, etc. the user has to speak carefully, extensively “train” the software and make many corrections when the software identifies the wrong word. This is not so bad when the speaker reviews the text afterwards, but disastrous for a “real time” conversation in a difficult environment! I think that a possible way round this would be to break the speaker’s voice down into phonetic information rather than text – but perhaps someone knows better?

Cheers, Jim Moritz

73 de M0BMU