Dennis Lee Bieber  wrote:
> On Sat, 17 Mar 2018 20:12:48 +0000, druck  declaimed the
> following:
>
>
>> Speech synthesis, as opposed to speech recognition, is not very
>> processor intensive. The synthetic ones require the most computation,
>> and the natural voice ones are more dependent on memory bandwidth and
>> storage latency.
>>
>> We were able to run many synths on 200MHz iPaq StrongARM PDAs and had
>> all the big name synthetic and natural voice synthesisers running on
>> easily on early 400MHz XScale Windows CE mobile phones. Even a Pi1
>> wouldn't have problems running those, and I suspect should also be able
>> to cope with the latest versions.
>>
>
>  The 8MHz Commodore Amiga used to have translator (converted normal text
> to encoded phonemes [numeric intonation data and conversion of "c" to "s"
> or "k" as appropriate) and narrator device. However, I think they lost the
> distribution license for the libraries by the time of AmigaOS 3. The
> synthesizer device had the ability to return height and width data to a
> running program, intended, I'm sure, to allow the program to animate a
> mouth to match the syllables.

Well, since we’re exploring the performance threshold for software speech
synthesis, the 1MHz Apple II ran SAM (Software Automatic Mouth), which in
its first incarnation drove a speaker from an 8-bit DAC, but in later
versions was fully software, using ultrasonic PWM to output quite
acceptable speech through the 1-bit built-in speaker interface.

Virtually all vintage text-to-speech synthesizers relied upon a
text-to-phoneme algorithm created by the Naval Postgraduate School in
Monterey, California. It worked reasonably well, allowed for pitch and
other “hints”, and was public domain.

As has been noted, synthesis is relatively easy; analysis is hard. ;-)

--
-michael - NadaNet 3.1 and AppleCrate II:  http://michaeljmahon.com

--- SoupGate-Win32 v1.05