Slashdot Mirror

We also offer quite a large range of languages. Our Canadian French voice, which was just released, is fantastic! Looks like marketing hasn't put him on the demo page yet though... :(

Todd

We also offer quite a large range of languages. Our Canadian French voice, which was just released, is fantastic! Looks like marketing hasn't put him on the demo page yet though... :(

Todd

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.

SPWX product listing

and yes, I do work there. Todd

That system was done by (SpeechWorks.

Thanks for the compliment!

Todd

(Full disclosure: I have worked with most of these companies).

Telephony-based voice-recognition is going to be the Next Great Thing (tm). The main companies that are involved in this stuff are SpeechWorks, Nuance (both work on the main speech recognition/software stuff), HeyAnita (which works with Sprint), and TellMe.

OpenVXI 2.0 was released just last week. According to the message on the VXI-discuss mailing list:

OpenVXI is a portable open source library that interprets the VoiceXML 1.0 dialog markup language. It provides a full implementation of the VoiceXML 1.0 specification, including all required features and nearly all optional features. Where the VoiceXML 1.0 specification is vague or incomplete, OpenVXI follows industry direction to fill the gaps.

See http://www.speech.cs.cmu.edu:/openvxi/ for details and source and binary downloads.

There is currently support for Windows (binaries are included) and Linux. Developers are currently working to add Solaris and Mac OS X.

NOTE: This is a VoiceXML interpreter. A real system would require a full speech recognition engine and a full text-to-speech implementation. SpeechWorks International ships a commercial version which connects to their recognizer and TTS products. This is a good playground for experimentation.

OpenVXI 2.0 was released just last week. According to the message on the VXI-discuss mailing list:

OpenVXI is a portable open source library that interprets the VoiceXML 1.0 dialog markup language. It provides a full implementation of the VoiceXML 1.0 specification, including all required features and nearly all optional features. Where the VoiceXML 1.0 specification is vague or incomplete, OpenVXI follows industry direction to fill the gaps.

See http://www.speech.cs.cmu.edu:/openvxi/ for details and source and binary downloads.

There is currently support for Windows (binaries are included) and Linux. Developers are currently working to add Solaris and Mac OS X.

NOTE: This is a VoiceXML interpreter. A real system would require a full speech recognition engine and a full text-to-speech implementation. SpeechWorks International ships a commercial version which connects to their recognizer and TTS products. This is a good playground for experimentation.

SpeechWorks is a commecial company bringing AT&T Labs's TTS stuff to the marketplace. They're currently offering Speechify, which is a slightly older version of the AT&T Labs system. A demo is here. Can anyone tell the difference?

SpeechWorks is a commecial company bringing AT&T Labs's TTS stuff to the marketplace. They're currently offering Speechify, which is a slightly older version of the AT&T Labs system. A demo is here. Can anyone tell the difference?

For what it's worth, SpeechWorks International licensed an earlier version of the AT&T synthesizer. You can find demos here. The version in the NYT seems to have been developed with different constraints. Many TTS engines are designed to achieve real time play back or to use limited amounts of CPU. For instance, synthesized speech during game play should only use 5% or maybe 10% of the processor. Whereas a system for Hollywood may demand considerable CPU power to produce small utterances (say 100 CPU seconds per second of speech). This is completely acceptable for many purposes where perceived quality is the primary criteria.

There is also an open source TTS engine called Festival, developed at the University of Edinburgh and at Carnegie Mellon University. You can find out more here. Or, just download the source.

[Disclaimer: In the interest of full disclosure, I work in speech recognition for SpeechWorks International. You may enjoy testing the SpeechWorks Demo Line at 1.888.SAY.DEMO (1.888.729.3366).]

A number of users have posted comments questioning the benefit of speech recognition operating on a PDA. Before examining this topic, I would like to quickly review the technical state of the art (in an effort to satify your inner-geek).

The Technology

Speech recognition operates by performing statistical matches of incoming sound against familiar words or phonemes (i.e. individual sounds; a works is composed of one or more phonemes). Traditionally, embedded speech recognition systems have featured small vocabularies (i.e. a limited set of recognizable phrases), but advances in processor speed are allowing larger and more complex vocabularies.

For applications in the telephony industry, a system running on a 500 MHz Pentium may support up to 50 lines and ten languages. These systems use a combination of directed dialog (asking for specific pieces of information - "what is your account number?") or limited natural language ("I'd like to fly from Boston to San Jose").

Returning to the PDA market, the task is to recognize a single user operating in a single language. This greatly reduces the memory and processor requirements. Further tradeoffs are possibly by adapting to the speech patterns of a single user as frequently occurs in dictation systems. But, as we will discuss below, the vocabularies are much more complex and word-spotting becomes vital.

Speech on a PDA

For simple tasks like navigation, the point and click interface works great. You get immediate feedback from the screen and you may peruse a page of information at a time. A speech based interface, in contrast, is more serial than parallel. If you are walking through a list of 50 items, your eyes will locate the correct item far faster than if the list is being read. Likewise speech recognition will not replace the keyboard for data entry. It is, however, a valuable suppliment which allows the user to jump to information not readily visible. While you're composing an email on the Palm Pilot, for instance, saying "Tell me the birthday of Jim Bob Jones" may be faster than navigating there yourself. Likewise, if you're navigating through a database of 20k companies, it may be easier to just say "Yoyodyne Propulsion Systems".

To make speech recognition useful on a PDA, the vocabularies must directly relate to the installed applications and information. Complex navigation using true natural language is a difficult and very much unsolved recognition task. But speech recognition on a PDA is even harder. Why?

Imaging that you're sitting in a cave and you hear "Dave, I'm sure that I've got it.. umm... that's not... no... Boston Sand & Gravel... come on...". You're the PDA. What does the user want? If you understood the context of the situation, you might recall the above example of company names in a database. You might say, I've got that installed and locate the entry for the 'Boston Sand & Gravel Company' for the user. But a PDA is not that smart. It needs to first pick out the allowed phrases from the noise and surrounding conversation. This is called 'word spotting'. Then it needs to decide how to interpret the phrase. Without a restricted application, the PDA must understand the context, frequently in human terms, of the speech.

If this seems hopeless with today's technology, you are correct. We will see speech applied first to limited interactions and simple applications. Over time, the domain will grow. Think back to handwriting recognition on the early Newtons. We've come a long way in a few years. On the PDA, the same will be true for speech.