Slashdot Mirror
State of Speech Synthesis and Text-To-Speech?
Gnulix asks: "Are there any, preferably either open source products available that produce realistic speech from an arbitrary (English) text? Projects such as Festival doesn't sound all that much better than SAM (Software Automatic Mouth) did on a Commodore 64 back in 1979, nor does SoftVoice's or IBM's new products sound very good. I mean we all know that Stephen Hawking is a fun loving guy, but I bet you that he didn't choose his unrealistic, robotic voice just for the heck of it. With all the amazing advances we have seen in real-time graphics, shouldn't speech synthesis have come much, much further than what is, seemingly, available today?" Ask Slashdot last handled the Voice-To-Text issue in January of this year.
is the best Text to speech conversion program
checkout http://www.naturalvoices.att.com/
Actually, I heard that they offered Hawking a revamped speech synthesizer, since although his was state-of-the-art in the seventies, today we have much better. He declined, saying he and his friends had gotten used to the voice, and it was "his". In fact, whenever on hears that particular flavor of voice synthesis, it's difficult not to think of Hawking.
He does relate, however, in A Brief History of Time, that at first people had trouble understanding "his voice", so that when he would speak or answer questions at lectures, he would have an interpreter who was more familiar with his voice repeat what he just said.
Interesting stuff...
Natoinal Weather Service describes their new system.
Yeah, closed source :)
MS has had text-to-speech as a object you can embed in your program with one line of VB code (same as you can embed IE) for a while now.
Apple has had text to speech entensions in tons of different voices for a long time. Some of the G4s used to read dialog boxes to you by default if you didn't click on them fast enough. Pretty unnerving the first couple times.
Several voice activated automated attendant systems I have called for my credit card and bank are amazing these days. They have insanely accurate speech recognition and really good text-to-speech.
So I wouldn't say the field is not advancing... it is.
Of course, a Google search for "open source text to speech" without quotes yields many promising looking hits, which I havn't evaluated. Why didn't you search there before asking Slashdot?
I've had enough abrasive sigs. Kittens are cute and fuzzy.
Nor, to harp on my pet peeve, do we have a theory of semantics that can put XML to any important use on the average webpage. These all need a model of the human psyche, because all human language is flavored with metaphors from the realm of motives and plans, etc (the psychological realm). Psychological science isn't delivering the sorts of models that NLP-etc need, and probably won't for many decades yet. [My AI FAQ]
And the best I have found so far is Festival with Mbrola voices (although not perfect they are far superior than the Festival voices)
For voice control stuff I found a little program called cvoicecontrol to be quite nice.
Hitler's in the fridge.
Just check THIS out:
http://www.naturalvoices.com/
quite a big step in the right direction in my opinion.
With all the amazing advances we have seen in real-time graphics, shouldn't speech synthesis have come much, much further than what is, seemingly, available today?
We haven't had that many amazing advances in graphics. Natural speech is to advanced raytracing what current text to speech is to current graphics. We still cannot raytrace in a single system in real time at the resolution of our eyes, and we still cannot produce natural speech in a single system in real time at the resolution of our ears.
Furthermore, we know less about the math of speech than we know about the math of light. Go visit your local university that has a good CS program, and browse the bookstore for the books used to teach speech recognition. In that book you will find that the average sound a human makes goes from production of complex, multitonal sound from the vocal cords through as many as five complex natural filters (body cavities between the vocal cords and lips) before it reaches the ears of the recipient.
Modeling these filters for one sound is hard enough. Each letter in our alphabet, except simple vowels, changes the filters throughout the letter. Furthermore the filters for a given letter may also change depending on the previous and next letter.
A system to create speech, therefore, must generate hundreds (perhaps thousands) of different filtered 'noises' just to reproduce the english language. Other languages can be much more complex.
Current common technology is to simply record the hundreds of 'simple' sounds and add them together. Really good programs use hundreds of hours of speech by voice actors to get several hundred sounds.
The penultimate is to mathematically recreate every part of the human vocal system from the lungs to the lips. This has obviously not occured. The computers may well be powerful enough, but the understanding of the vocal tract is extremely limited.
In other words, wait 5-10 years. There still isn't a killer application for text to speech, but with devices getting smaller and smaller, there will be soon enough.
-Adam
Another extremely strong competetor to Natural Voices is Speechwork's Speechify. Take the "Speechify Challenge" -- it's still possible to tell which is a real recording and which is the computer, but it is very difficult. Some say it's the best engine available, but I guess that's a matter of personal preference.
I don't know about Open Source TTS, but the commercial versions (AT&T, Speechworks, and others) are sitting on the threshold of truly natural speech. I work in the speech industry, so I follow progress and have seen some of the unreleased demos of upcoming versions. In the next couple years, we can expect amazing things. It won't be long before the Speechify Challenge will truly be impossible to beat.
By the way, for those of you who don't know, the newest and best-sounding engines don't use purely synthesized sounds as older and small-footprint engines do (Festival and Steven Hawking). The engines are built using actual recordings: a "voice actor" will sit in a studio and record dozens of hours of speech, and then, over the course of several months, the recordings are then cut and spliced into individual phonyms, which are reassembled by the engine. This means that the voices actually sound like real people, and the only unrealistic part is the inflection when generating complete sentences. You can order custom voices (for several tens of thousands of dollars) and get a voice that sounds identical to that of your celebrity of choice.
One addendum: the fact that the newest engines use real recordings is exactly the reason why it will be nearly impossible for Open Source engines to approach the quality of commercial versions. The amount of work involved in extracting the raw sounds from recordings is staggering, and it requires full-time commitment from trained experts over the course of many months (not to mention the cost of hiring voice talent). There is no way to avoid the costs involved, and so Open Source alternatives cannot become available without some sort of large grant. Unfortunate.
Here at work we monitor progress of and/or use the following:
DECTalk (One of the most widely used)
Eloquent (http://www.eloq.com - dead URL?) (fairly natural-sounding with dialects)
Elan (European languages)
They've all been improving over the years.
- dyphone-based synthesis where the database contains one dyphone (end of first sound + start of next sound) for each psossible sound combination. This approach is used in Festival. Dyphone-based synthesis will hardly sound better that in Festival because dyphones have to be modified artificially to fit every variation of pitch, duration and any other parameter that is needed to produce a given phrase.
- corpus-based synthesis takes a different approach where a large database of several hours of speech is recorded and manually labelled to mark the start and end of each sound. Such a database is used to extract the best and the longest sequence of dyphones during the production. This approach gives naturally sounding results for short sentences where intonation is not so important
Given that the cost of developing a database for corpus synthesis may easily be 100 times higher than for dyphone synthesis, there are very few companies that make them. Two companies offer a demo on the internet: ATT and Scansoft (former L&H) and