I've also noticed this stagnation. The quality of the spoken voice TTS sound depends upon two things, I've heard -- processor speed and memory (RAM). Processor speed has increased dramatically in the past few years. I wish someone would design TTS that only works on the fastest processors. There seems to be too much lowest-common-denominator going on in this field.
I was under the impression the problem the field currently has is that improving the diction rapidly becomes AI-complete. Voice synthesis is either a solved problem, or one that could be solved with just a bit more effort, but the real problem is what do you feed that perfect voice model?
Read your comment or my comment aloud to yourself. Now feed it to your choice of TTS engine. The problem isn't the words, the problem is the lack of comprehension.
I've thought about getting into this area myself, but I was too afraid there was not enough market for it. This was several years ago. As far as CPU power, today's average PC is easily 5x what's necessary for perfect speech. The real question is the algorithms being used. Perhaps the fear is that the algorithm would be pirated. I mean just look at the history of digital audio (or video) and encoding, with things like Xvid and Ogg. Basically every time a good algorithm even starts to gain traction, an open "alternative" is made available practically overnight. This is not to say I don't like open algorithms. In fact, I believe that any standard algorithm should be open. This fact, however, is enough to deter research in this field. Perhaps a Web service that converted text to speech would be one option, but it would have limited applicability.
Edit: Perhaps a Kickstarter or related would be a good idea since this type of feature would be useful by so many people. Nearly everyone has functioning ears. (no offence to those who don't)
Basically there is Nuance, which is the leader. All others either use older open source projects underneath, legacy software or just license from Nuance. If you can name another one, please do.
Do you mean the current execution is bad or that it is just a bad concept? I can think of plenty of times where TTS is the perfect interface for the problem at hand. For example, a Sat Nav giving turn by turn directions.
That is true - all voice related interfaces are best exemplified with travel. STT for example is best when you're alone and can't use your hands aka driving. TTS is best for the same set of situations, so listening to articles on the train or getting directions read back.
In general, TTS is a better interface than STT, both are (in my very, very humble opinion) bad interfaces.
What do you see as the business use case for more realistic text-to-speech?
We use TTS extensively within the Pantelligent iOS and Android apps, and it's something our users requested and get a lot of value out of. It seems like the existing solutions are already good enough / dramatically above the threshold of usefulness for an interactive real-time-guidance application like ours, and just keep getting better from the OS side.
It seems that IVR phone systems could benefit. Maybe the state-of-the-art just hasn't had the chance to propagate, but it always strikes me as inelegant when I hear, "Your account has a balance of... ONE, HUN-DRED, DOLLARS... AND... NINETY, FIVE, CENTS. To pay this bill,"
I've just tried playing "Your account has a balance of 100 dollars and 95 cents" on https://www.ivona.com/us/ and it seems quite natural to me (but I'm not a native english speaker so I may not be sensitive).
- The "bank" one referenced above: this is made of short recordings of a real person saying the words or phrases, cut up and then concatenated. For some messages they sound exactly like a real person (because all it does is play a single recording), but when numbers are inserted, the above characterization is quite accurate. There is no effort to make the inflection fit properly in a sentence or have it sound natural.
- ivona.com and OS X `say`. These generate audio in real time, and may have a few samples but are generally created on-the-fly according to what is around the text. This is where the research is at right now, but the main problem is the CPU required to generate these. Your car, or Madden 2015, or the bank might not want to use up too much CPU time to make their audio sound like that.
In the late sixties I was programming at a bank and even then there was a voice response account-balance phone number. The machine was quite large, about the size of two refrigerators. It had twelve different loops of tape that spun around. Each tape had one digit or word. It sounded exactly as you describe.
It wouldn't surprise me if that were a minefield of liability, or if there was research that shows that doing it that way makes the message more intelligible.
Either way, the incentive for the bank to change things must be minute.
We build video content and localize it in multiple languages, we use voice artists to record the audio for this content, obviously this has scalability issues. We would use TTS if we could find an engine that sounded more like a human and less like a computer, but from our testing of various engines (including the ones posted in the comments) they don't come close to sounding like humans. We are just one business case in a huge sea of opportunity.
I've messed around with Api.ai (https://docs.api.ai/docs/tts) and it is quite impressive. Full-featured, amazing pricing (free), and goes way beyond just TTS.
> It seems like TTS technology hasn't evolved much over the years and I was wondering if any startups are working on making it sound more realistic?
It's not true that TTS hasn't improved (see below). Many people are working on this, both in academia and in private enterprise. It's an obvious and potentially valuable part of the human-computer interface.
This is not to suggest that it's easy -- the mathematics and vocal tract modeling problems are formidable. The only reason there are reasonable TTS resources now is because of the rapid increase in computer power -- power that's needed to support this feature.
Here's a site chosen at random that offers a high-quality TTS example:
OSX's "say" is amazing (judging from their english and german voices). Do you believe there's enough room for improvement to build a business case on it, even though some companies have worked on the problem for decades?
Ivona, acquired by Amazon, is a hugely notable example. They're great folks, and do great work. Alexa'a voice was done by them, but an early customer of theirs was the Polish public transit system.
A lot of really great work is happening in academia; I'm not going to name names because I'd forget someone deserving.
(Shameless plug: we [0] do speech and language consulting including custom TTS.)
aah. Okay. Disclaimer, [b] this isn't my area [/b] and I'm going to miss some folks. I apologize to all of the important and interesting work that I'm missing.
Also, I'm really linking to research groups, so take these names as starting points and look at their students and other professors working with them.
First off, the Blizzard challenge is a major hub of activity. [4] Festival is an important piece of software [5]. Interspeech is an important conference [6] (take a look at the speech synthesis track and the organizers for that).
Alan Black [0] @ CMU is kinda a giant.
Keiichi Tokuda [3] @ Nagoya also a giant.
Simon King [1] @ Edinburgh, just had a paper linked to on HN a few days ago and does important work.
TTS has definitely evolved over the years. If you compare Google Maps voice to Stephen Hawking, it's night and day.
However, I can definitely understand how TTS technology looks stagnant. Part of this is that going from nothing to something reasonable happened exceedingly quickly. Early TTS research was supported by the US government which saw that early systems were comprehensible (if not wonderful sounding) and declared victory. Funding went to other problems in computational linguistics (like speech recognition, information extraction, etc.) and so did a lot of the workforce.
Modern systems usually involve many hours of speech from a single person and use variable length units to form more natural speech. Many systems still sound pasted together because that's how enterprise technology goes. How many banks have online banking that seems like it's from the 90s? You can't compare what systems can do to what some call center has installed as its technology. Someone has linked to Inova. Google and Nuance have good systems as well, but there's a balance between resources and perfect speech.
In terms of some of the issues. . . When you're going through a finite amount of recorded speech, you have to choose something that fits. It isn't going to be perfect in many cases. You have to deal with things like F0 declination. You have to deal with how long phonemes are going to be. You have to deal with breaks in utterances.
And the fact is that we can understand Hawking's 1980s TTS.
If you want to start thinking about the problem more, try inputting these two statements into Inova "Do you really want to see all of it? Do you want to see all of it? I want to see all of it." Notice how it tries to rise around "really" in the first sentence. It's trying to match how we would speak - rising for "really" in the first sentence and rising for the question-ending in both questions. But it kinda misses in both cases. Still, in some ways it's amazing that it recognizes "really" as something that should go up. It recognizes that questions go up at the end. It recognizes how the non-question goes down as the sentence progresses. And it finds things within its data set to fit to how it thinks the sentence is going to go. But it doesn't have perfect language understanding so it doesn't know exactly how things would be said - a lot of sounding natural isn't making the phonemes more accurately, but the intonation and attitude of the speech. It also has to find something that fits. Lots of smart things are done, but it's pulling from a limited amount of recorded speech - speech that has been sliced in many useful ways, but still limited.
TTS has definitely evolved and I think that Google, Nuance, and others are definitely pushing it forward. You're going to interact with a lot of legacy systems that feel like they're still in the Hawking era. But most ATMs I use don't even have touch screens (opting for buttons on the side of the screen) and even fancy ATMs like Wells Fargo don't feel like an iPad. You don't want to compare to systems that are so far away from modern, commercially available systems.
There is definitely work being done on it and it's definitely become much better. But to an extent, it isn't something that a lot of companies are going to work on. How big is the market for TTS? Before you say, "it's useful in loads of things," think about the market for maps. A lot of it is Google or Apple Maps. Loads of apps integrate mapping, but don't want to map the world or run their own infrastructure for serving it. Some use OpenStreetMaps, but they're really just serving generated tiles rather than re-mapping the world. If you were to create a TTS startup, what would your business model be? Pay us money to TTS your text rather than getting it for free from Google's Android TTS? The issue is that TTS is more a feature t...
A ML based app to read articles to me in the morning on my way to work may have some commercial success. The ones currently have a hard time reading through an entire article without sounding like an robot, or reading a headline as if it was part of the previous sentence.
I made this app called Ultimate Alerts that allows for all email and text messages to be read over TTS when your car goes over 10mph for over 1 minute.
It switches back to normal settings when you go below 10mph for over 1 minute. Helpful with switching everything to TTS when you're driving automatically.
43 comments
[ 5.0 ms ] story [ 98.8 ms ] threadRead your comment or my comment aloud to yourself. Now feed it to your choice of TTS engine. The problem isn't the words, the problem is the lack of comprehension.
Edit: Perhaps a Kickstarter or related would be a good idea since this type of feature would be useful by so many people. Nearly everyone has functioning ears. (no offence to those who don't)
1. Nuance
2. AT&T
3. IBM Watson
In general, TTS is a better interface than STT, both are (in my very, very humble opinion) bad interfaces.
We use TTS extensively within the Pantelligent iOS and Android apps, and it's something our users requested and get a lot of value out of. It seems like the existing solutions are already good enough / dramatically above the threshold of usefulness for an interactive real-time-guidance application like ours, and just keep getting better from the OS side.
- The "bank" one referenced above: this is made of short recordings of a real person saying the words or phrases, cut up and then concatenated. For some messages they sound exactly like a real person (because all it does is play a single recording), but when numbers are inserted, the above characterization is quite accurate. There is no effort to make the inflection fit properly in a sentence or have it sound natural.
- ivona.com and OS X `say`. These generate audio in real time, and may have a few samples but are generally created on-the-fly according to what is around the text. This is where the research is at right now, but the main problem is the CPU required to generate these. Your car, or Madden 2015, or the bank might not want to use up too much CPU time to make their audio sound like that.
Either way, the incentive for the bank to change things must be minute.
It's not true that TTS hasn't improved (see below). Many people are working on this, both in academia and in private enterprise. It's an obvious and potentially valuable part of the human-computer interface.
This is not to suggest that it's easy -- the mathematics and vocal tract modeling problems are formidable. The only reason there are reasonable TTS resources now is because of the rapid increase in computer power -- power that's needed to support this feature.
Here's a site chosen at random that offers a high-quality TTS example:
https://www.ivona.com/
It's pretty good based on prevailing standards, and it's the outcome of a lot of work.
To find the companies working on this, just Google for "high-quality tts".
I'm not even a stoner and folks seem to get a kick out of it whenever they're pair programming with me.
May be of interest to you.
A lot of really great work is happening in academia; I'm not going to name names because I'd forget someone deserving.
(Shameless plug: we [0] do speech and language consulting including custom TTS.)
[0] cobaltspeech.com
Also, I'm really linking to research groups, so take these names as starting points and look at their students and other professors working with them.
First off, the Blizzard challenge is a major hub of activity. [4] Festival is an important piece of software [5]. Interspeech is an important conference [6] (take a look at the speech synthesis track and the organizers for that).
Alan Black [0] @ CMU is kinda a giant.
Keiichi Tokuda [3] @ Nagoya also a giant.
Simon King [1] @ Edinburgh, just had a paper linked to on HN a few days ago and does important work.
Mark Gales [2] @ Cambridge does work here, too.
[0] https://www.cs.cmu.edu/~awb/ [1] http://www.cstr.ed.ac.uk/ssi/people/simonk.html [2] http://mi.eng.cam.ac.uk/~mjfg/ [3] http://www.sp.nitech.ac.jp/~tokuda/ [4] http://festvox.org/blizzard/ [5] http://www.cstr.ed.ac.uk/projects/festival/ [6] http://interspeech2015.org/wp-content/uploads/direct/INTERSP...
[1] http://www.vocaloid.com/en/
However, I can definitely understand how TTS technology looks stagnant. Part of this is that going from nothing to something reasonable happened exceedingly quickly. Early TTS research was supported by the US government which saw that early systems were comprehensible (if not wonderful sounding) and declared victory. Funding went to other problems in computational linguistics (like speech recognition, information extraction, etc.) and so did a lot of the workforce.
Modern systems usually involve many hours of speech from a single person and use variable length units to form more natural speech. Many systems still sound pasted together because that's how enterprise technology goes. How many banks have online banking that seems like it's from the 90s? You can't compare what systems can do to what some call center has installed as its technology. Someone has linked to Inova. Google and Nuance have good systems as well, but there's a balance between resources and perfect speech.
In terms of some of the issues. . . When you're going through a finite amount of recorded speech, you have to choose something that fits. It isn't going to be perfect in many cases. You have to deal with things like F0 declination. You have to deal with how long phonemes are going to be. You have to deal with breaks in utterances.
And the fact is that we can understand Hawking's 1980s TTS.
If you want to start thinking about the problem more, try inputting these two statements into Inova "Do you really want to see all of it? Do you want to see all of it? I want to see all of it." Notice how it tries to rise around "really" in the first sentence. It's trying to match how we would speak - rising for "really" in the first sentence and rising for the question-ending in both questions. But it kinda misses in both cases. Still, in some ways it's amazing that it recognizes "really" as something that should go up. It recognizes that questions go up at the end. It recognizes how the non-question goes down as the sentence progresses. And it finds things within its data set to fit to how it thinks the sentence is going to go. But it doesn't have perfect language understanding so it doesn't know exactly how things would be said - a lot of sounding natural isn't making the phonemes more accurately, but the intonation and attitude of the speech. It also has to find something that fits. Lots of smart things are done, but it's pulling from a limited amount of recorded speech - speech that has been sliced in many useful ways, but still limited.
TTS has definitely evolved and I think that Google, Nuance, and others are definitely pushing it forward. You're going to interact with a lot of legacy systems that feel like they're still in the Hawking era. But most ATMs I use don't even have touch screens (opting for buttons on the side of the screen) and even fancy ATMs like Wells Fargo don't feel like an iPad. You don't want to compare to systems that are so far away from modern, commercially available systems.
There is definitely work being done on it and it's definitely become much better. But to an extent, it isn't something that a lot of companies are going to work on. How big is the market for TTS? Before you say, "it's useful in loads of things," think about the market for maps. A lot of it is Google or Apple Maps. Loads of apps integrate mapping, but don't want to map the world or run their own infrastructure for serving it. Some use OpenStreetMaps, but they're really just serving generated tiles rather than re-mapping the world. If you were to create a TTS startup, what would your business model be? Pay us money to TTS your text rather than getting it for free from Google's Android TTS? The issue is that TTS is more a feature t...
Has anyone else wondered why Stephen hasn't upgraded his voice? Maybe it is his signature of sorts.
That's exactly what he's stated publicly - it's so widely recognised it's part of his identity.
It switches back to normal settings when you go below 10mph for over 1 minute. Helpful with switching everything to TTS when you're driving automatically.
Lots of other functionality as well. Check it out: https://play.google.com/store/apps/details?id=com.org.imsono...