Jay's blog on text-to-speech (Now Defunct)

Moving away from Text to Speech... but other great folks are still here

jaywaltm — Tue, 03 Oct 2006 01:16:00 GMT

I'm no longer blogging on text to speech; however, there are still a lot of great people blogging on speech at Microsoft. Your best contact is going to be Chuck Opperman who has a blog related to TTS and Speech API content:

http://blogs.msdn.com/chuckop/

Also, the Speech team participates on newsgroups and on a Yahoo Speech forum if you are looking for additional insight.

Best,

Jay

Model Talker will allow individuals to create their own custom TTS voice

jaywaltm — Wed, 12 Jul 2006 05:03:00 GMT

It was only a matter of time.... AgoraNet, Inc. appears to be in the process of introducing the Model Talker system of TTS, allowing an individual to create a TTS engine based on her voice to be used with any Sapi 5.0 compliant system (Windows XP/Vista). The technology is targeted towards individuals loosing their ability to speak. Very cool.

I can imagine that someday you'll check into a new company and as part of your new hire task list, you'll be ask to sit down at your desk for the first hour of each day recording a series of sentences. At the end of a few weeks, you'll upload your recordings to HR and they will run it through a tool and then send the result to your IT team. Now, when you send an email to someone in your company, they can listen to your email - with not just any TTS voice, but a TTS version of your very own voice.

The Blizzard Challenge: Rate text to speech samples

jaywaltm — Fri, 30 Jun 2006 19:43:00 GMT

What's the Blizzard Challenge? In a shell's nut, it's a competition that provides researchers and speech labs with the same set of originally recorded waves (i.e., 5 hours of a male speaker), and then challenges teams to create the "best" (i.e., most natural and intelligible) TTS voice. If it were an architecture challenge, it would be like giving the same set of building materials to many different architects to come up with the most functional and beautiful building.

And now rating the results is open to the public here:

http://www.speech.cs.cmu.edu/blizzard2006/register-R.html

It took me about 20 minutes to complete the study. There are five tasks, with the first three scoring waves, and a transcription for the last two. It's probably a good indication of how good a TTS system you can get with 5 hours of recordings. I can't tell who are all of the participant groups at this point. Festvox has a little bit more information on its site.

Vista Beta 2 Screen Reader and Text to Speech

jaywaltm — Fri, 09 Jun 2006 21:22:00 GMT

If you are running Vista Beta 2, check out the new text to speech functionality using the Narrator screen reader program. (Narrator is also enabled in Windows XP; however, the Vista TTS voice is way better in terms of naturalness and intelligibility IMO compared to MS Sam).

You can start Narrator in one of two ways.

1. If you have the "Run" command enabled in your start menu. then go to START > Run > and type "narrator.exe". (Egads, no, the 'run' command is not there by default. To get it there, right click on the toolbar and go to PROPERTIES > click on the START MENU tab > click on the Customize button and then find the check box for the "Run-command on".)

2. Or, simply go to the control panel and click on "Ease of Access". There you'll find the "Narrator" button to get you started.

Text to Speech in Mission Impossible 3: A Dissection

jaywaltm — Fri, 12 May 2006 23:47:00 GMT

Besides being the best of the three MI movies, there were 2 instances of TTS in the movie that deserve some discussion (and clarification). One of the scenes was simple and plausable, while the second was a definite stretch (i.e., not doable by today's technology).

In the first scene with TTS, one of the "good" guys was automating the descruction of a wharehouse full of "bad" guys, using vehicles equipped with large guns. When the automation started, the computer began speaking out some information using TTS. I'm pretty sure it was Mac OSX TTS. Definitely low on the naturalness scale, but intelligible nonetheless. (Can anyone confirm which TTS voice this was?)

In the second scene(s), THE "good" guy (i.e., Tom Cruise's character), forces THE bad guy to read several syntactically but not semantically grammatical sentences off of a business sized card at gun point. Within seconds of completing the reading of the card, another "good" guy has intercepted the wave beneath the complex to generate a highly natural and intelligible TTS voice which is sent back to our protagonist in a bathroom who then can talk with the "bad" guy's voice. OK, so I'm actually quite forgiving in movies, giving the technology the benefit of the doubt (i.e., I pretend that I'm watching Sci-Fi and not a modern day action movie). So, if we assume that this was some other technology beyond TTS, great. No worries. However, if you are insisting that the movie follow current plausable technology, then here's what wrong with the TTS in this second scene:

1) The TTS engine was generated from several sentences. Today, takes many many hours of recordings to generate a naturally sounding engine.

2) The recording was done in a bathroom next to a loud party and then streamed to a nearby underground location. Not likely to result in the high quality recordings that one would need for TTS.

3) The recording was streamed through rock. I'm imagining that some signal loss would be encountered in real life.

4) The resulting TTS sounded almost EXACTLY (egads, as if it truly was the other actor speaking with Tom lip-synching) like the "Bad" guy! Even on the BEST concatenative engines (i.e., based on 40+ hours of recording a person's voice), it won't sound just like the real person.

Comments? Alternative takes?

Problems with French Liason in TTS - "A thread is a happy husband?"

jaywaltm — Thu, 23 Mar 2006 23:49:00 GMT

Recently I was talking to a colleague who is a native French speaker (hey, that's the great benefit of working here at Microsoft - lots of non-English speakers around) and she pointed out an interesting occurance of mispronunciation in French TTS that she found while casually taking advantage of speech technology (she knows I dig TTS). The phrase is the following:

"Un fils et un mari heureux"

Literally, this translates in English as "A son and a happy husband"; however, it comes out sounding like "A thread is a happy husband!"

Why is that? Well, the word for "son" should be pronounced as [f i s] but comes out as [f i l]. And "fil" means string or thread. Also, the "et un" string should be [eh ah] as in "and a", but comes out incorrectly as [e tah] meaning "is a." (please forgive my phonetic convention). All you linguists know that this is all due to liason.

TTS is cool... but only as good as its sons, threads, and husbands.

Shakespeare video with text to speech

jaywaltm — Wed, 15 Feb 2006 00:15:00 GMT

Egads, it's quite clear that Shakespeare wasn't meant to be read with a text to speech voice... or at least not with the Micorosoft Mary and Microsoft Mike which are representative of older MS technology (old in terms of what users will get with Vista). At any rate, check out this video if you want to see TTS acting at its worst.

This video suggests that having emotive TTS might be a feature worth supporting. I know that Loquendo has such a feature with its TTS, but it appears to be limited to a few pre-recorded emotive sayings.

What is TTS? How does it work?

jaywaltm — Wed, 08 Feb 2006 05:21:00 GMT

I recently went over the entry in Wikipedia on text-to-speech, and it turns out to be a really good overview of the technolgogy. If you are interested in its history and the various components that would have to be handled by any TTS system, I highly recommend this site. There are also a number of helpful links at the end of the article.

Another blogger on TTS

jaywaltm — Mon, 30 Jan 2006 22:23:00 GMT

Check out Blake's blog here and you can see some of the enthusiasm around text to speech. Blake's site, The Road to Know Where, is loaded with all sorts of tips for Windows users. I'll know where to send my mom the next time she is looking for helpful user hints. :)

Get El Mundo in podcast in Spanish

jaywaltm — Fri, 27 Jan 2006 06:25:00 GMT

El Mundo is one of the more well known newspapers in Spain. And they are now offering an MP3 podcast download of the daily news. It appears that the voice of Rosa is the very same Rosa of AT&T Natural Voices.

I found it interesting that the podcast actually caveats its TTSness on start-up. That is, it says, many may be surprised by my tone of voice... but to keep in mind that I'm a machine. My guess is that people are a lot less critical on a TTS voice if you tell them ahead of time that it's a computer rather than a real voice. That would be an interesting study if it hasn't already been done.

(Special thanks to Enrique Lopez Diaz for pointing out the El Mundo TTS site.)

Who else blogs on speech at Microsoft?

jaywaltm — Fri, 20 Jan 2006 00:45:00 GMT

(No, it can't just be me... oh, and it isn't...) Stephen Potter has an extensive list of bloggers across several teams that get a trill out talking Speech. And while I'm the only one focused on TTS (life beyond TTS?), a lot of interest around here is in Speech Recognition.

How do you pronounce the digit "1"?

jaywaltm — Tue, 17 Jan 2006 20:38:00 GMT

This is an interesting, if not challenging problem for TTS systems. So, how should you pronounce "1"? Should an English TTS system simply say "one" when it encounters "1"? (Okay, the answer is "no", and here's why:)

Let's look at the following English sentences:

(a) I have 1 friend. ("i have one friend")

(b) Can you meet me on 1/3/05? ("can you meet me on january third two thousand and five?")

(c) My birthday is on March 1. ("my birthday is on march first.")

In sentence (a), we can read the digit "1" as "one". But in sentence (b), the digit "1" can be spoken as "January" because it is in the context of a date. And in sentence (c), the "1" commonly takes on an ordinal reading by being pronounced as "first".

This problem of properly disambiguating written text into correct word expansions is known as text normalization. For TTS systems, it's very problematic depending on the context. While digits are most often affected in terms of frequency of occurance, other kinds of patterns are also very problematic (I'll save this for another post).

So, the digit "1" has a couple of pronounciations as just seen in English. But what about in another language such as Spanish?

Let's look at similar Spanish sentences:

(d) Yo tengo 1 amigo. ("i have one friend") - "1" pronounced as "un"

(e) Yo tengo 1 amiga. ("i have one friend") - "1" pronounced as "una"

(f) Yo tengo 1. ("i have one") - "1" pronounced as "un"

In these examples, the "1" can take on three different pronounciations! It's not the semantic context (e.g., "date", "time", "fraction") that requires disambiguation, but rather, the context is the gender of the word that the "1" modifies - or the part of speech of the "1". So, in sentence (d) the pronunciation is "un" because the following noun is masculine, but in sentence (e) the pronunciation is "una" because "amiga" is feminine. But, in (f), the pronunciation is "uno" because the "1" is itself acting as a noun.

Can you see why proper identification of part of speech is so important for text normalization?

You might be thinking of how to write a rule to capture the distinctions in sentences (d), (e), and (f). (For example, if there is a noun following the digit "1" then read the "1" as "una" or "un", otherwise, read the "1" as "uno".); however, can your rule account for the following long distance dependency?

(g) De todas las chicas en la clase, hay 1 que me gusta.

Does your TTS system get (g) correct? (Note, the "1" should be readout as "una".)

Listen to emails and word documents in MP3 format on Pocket PC

jaywaltm — Thu, 05 Jan 2006 01:44:00 GMT

Just when you thought you might be able to use your car commute time to sit back and relax, listen to NPR, and creep slowly across the Seattle bridge... TTS is there to readout that document that you didn't have time to read the night before.

Magnetictime has a product that I have not tested, but I thought I'd mention it here as TTS is core to its functionality. It also uses Acapella engines. "MT1" is the name of the program and apparently it will readout your outlook mail as well as word docs on your pocket PC device. By creating a new folder in Outlook, text is converted into waves and then the program can transfer the waves directly to your device. Currently works for English, but soon to work with more languages.

Seems like a cool app if you have a device that's not a smart phone. It doesn't appear that you'd get up to the minute info... you have to download to the device before you hit the road. And at least you aren't using precious airtime minutes to use it either.

Deskbot adds cool animations with text to speech

jaywaltm — Fri, 30 Dec 2005 00:35:00 GMT

I recently found an article describing the Deskbot which is one of the cooler text to speech reader programs available IMHO. Oh, and it's free! (BTW, does your TTS engine correctly readout the string "IMHO" as "in my humble opinion"? -- it should). (BTW, Neospeech's Paul reads it right, but XP Sam does not).

So what's Deskbot do? After it's installed, a small animated character appears on the screen (you can put the character anywhere you want), and then allows you to speak text in any window when it's pasted to the clipboard or when the the character is double clicked when there's highlighted text. It will also announce the time on specified intervals. And the best part is that you don't need to have loaded some special screen reader program.

You can also change options like altering the voice (pitch, speed, volume). Oh, and you can even change the character by downloading an MS Agent compatible character through the MS Agent Ring site.

New Hindi text to speech voice. Are you a native Hindi speaker? What's your opinion?

jaywaltm — Wed, 14 Dec 2005 00:18:00 GMT

There was an ariticle on web that describes a company with a TTS voice for Hindi. And for all you SAPI devs, the desktop version is SAPI compliant. According the article, "'Vaachak', developed by Prologix Software Solutions, is the first, high-quality, Indian language TTS system, which converts Hindi and English text into natural and pleasant sounding speech."

So, is it any good? I can tell you that the acoustic quality, at least from the demos, sounds a bit compressed and there is certainly some artifacts from concatenation. But still, who else out there has a Hindi TTS system?

What do you think? Are you a native speaker of Hindi? I'd love to get your impressions of the sample waves or you can use the demo to enter in Hindi text and have it read out.