Hacker News Reader

Comment by apendleton | Hacker News Reader

It's a shame that the quality of open-source text to speech engines is so much worse than the current commercial state of the art, as that's the most notable difference in the demo videos between this and something like Siri. Would fixing that just be a matter of recording more high-quality free sample libraries, etc., or are there fundamental technical challenges to solve?

View on HN

It's mostly a question of training data. Google trains its acoustic models on thousands of hours of annotated audio samples. It's very hard of an open source project to i/ get enough data ii/ have the computing power to actually train with it.

> ii/ have the computing power to actually train with it.

That's hardly an issue.. You just have to know one person that works for a university and you are pretty much all set.

I wonder if the audio data from television programmes combined with subtitles would make a good training set. Here in the UK many of the subtitles are very good quality (presumably human transcribed), and would be easy to strip out non-language noises such as laughter (normally transcribed in brackets).

Google used the GOOG-411 automated directory assistance number to collect data for training.

I don't think that's true.

That is a big factor in the accuracy of speech-to-text engines, but I don't think training data is a problem for text-to-speech (which is what the OP was complaining about).

I strongly disagree! Text to speech needs the same data as speech to text - a well annotated collection of raw, single speaker speech data from a variety of speakers and accompanying text labels.

It is very, very difficult to find a large, well curated dataset of speech with accompanying text labels. TIMIT is the gold standard of speech recognition despite its simplicity, and it costs ~$150(!) to get access at all. Switchboard is another famous one which is bigger, but only partially labeled and still very hard to get. CMU has an openly available, small dataset called AN4, but it is only 64MB of raw recordings (a lot for 1991 when it was recorded, but still small today).

You could use YouTube and their automated translations, but the text is shoddy at best and definitely too unreliable to feed to an algorithm without significant cleansing.

Honestly, U.S. Congressional speeches may be the best "open" option IMO - good transcriptions, a variety of speakers, and TONS of data. If you could sync up the transcriptions with recordings of C-SPAN it could work very well, though I have not seen a dataset that puts this together. This still only covers English!

Companies understand that data is valuable - that is why they take the rights to it in every EULA! Well curated internal datasets will be the KFC/Coke secret recipes of the "big data" age.

I recently discovered vocalid.org thanks to a TED video[1]. They work to provide people having speech impairments with unique synthetic voices thanks to voice donors. According to their process, a voice donor need to read 2-3 hours of text to allow them to build a custom synthetic voice.

I don't know if they have any plan to open their data but if someone could find a way to increase the amount of voice donors while improving an open dataset, it would be interesting. Just a thought, they are most certainly many more issues / options in this field.

[1]http://www.ted.com/talks/rupal_patel_synthetic_voices_as_uni...

I can't upvote this enough. There are very few companies that are masters of this game. Because it is extremely difficult to build. Every stage is a painful process and there are so many moving parts. To begin with you need very high quality recordings. Successful companies have voice directors whose job is to record voice at most optimal settings (and to make sure speaker speaks is neutral dispassionate tone). And then there is the manual process of segmentation and phoneme alignment - That requires lot of human effort. Then you run the code! It will take quite sometime. Then you study the quality of synthesis. Fix issues manually (for example, if you are using concatenative synthesis, you may have to fix offending phoneme alignment based on feedback). And iterate on this. Unless you have a team, I wouldn't venture into this and rather use third party leaders.

finally a use for the filibuster .

Sounds like this field could use some crowd sourcing.

What kind of data would be the most useful for learning? Would it be the same phrases read by different people, single words or is any text good? Do we need lots recordings from a single person or many smaller samples from different persons?

I'm thinking of a web site where people could contribute to the project by joining and then reading out phrases the system shows them.

You mean something like http://www.voxforge.org/

Seriously there is no massive CC-licensed source of audio data out there. Most of the fancy algorithms for doing speech recognition are on github. What isn't is a massive and diverse dataset. I encourage others to reply if they have seen otherwise.

Economics is the main challenge because it's a difficult task and the demand for it is low (market is small if demanding [0]). But the mbrola people, who have a nice free speech synthesis program that also isn't opensource, [1] say it's a hard technical problem too.

[0] http://www.ted.com/talks/roger_ebert_remaking_my_voice

[1] http://tcts.fpms.ac.be/synthesis/mbrola.html

This is using espeak. There are alternatives, such as Festival which, with decent voices, sound way better.

Example: https://www.youtube.com/watch?v=BA0z6ztG7qU

The question is: how do you make it sound that good? Is that with commercial voices?

Despite people mention the lack of the data it's not a problem anymore. It was a problem like few years ago but these days we have access to thousands hours of speaker data. For example we have 100 hours of single speaker database at CMUSphinx, it's not a big problem.

Algorithms become a bigger problem these days. For TTS for example one has to implement hybrid speech synthesis technology combining hidden markov models and unit selection. Most commercial companies are using this technology, but no open source project has it released. The closest ones are OpenMary and Festival, but they are either unit selection or hmm, no hybrid synthesis implementation yet.

A lot of the DEC Voice stuff was part of a DARPA contract, it may be possible to liberate from some archive somewhere.

You can use Google TTS. It doesn't have a documented public API, but nor did Google Reader.

It's a shame no one did something you don't even have idea how to do? Do you really think it's a shame?

I'm normally not a fan of condescending comments, but you hit the nail right on the head. It doesn't seem like some people grasp that a lot of open source projects are coded by unpaid volunteers in their free time in subjects they're passionate about.

Granted, there are commercial open source projects.

I'm sure everyone on HN is familiar with how open source projects work. The poster wasn't trying to tell contributors to get their act together and produce a higher quality product.

Beyond just lamenting the quality gap, they were trying to understand what it was that open source projects of this nature are missing. Is it lack of research making its way downstream? Poor media quality? There was no blame placed on contributors.

Understanding the answers to these questions can even help someone like the parent find an 'in' to contribute something themselves.

Personally, I really think it's a shame, yes. Why wouldn't it be?

Care to check a dictionary first? apendleton is not saying open source programmers should be ashamed that they haven't produce a high quality speech engine. Would you feel better if it was phrase like "it's a bummer that.."? Less hurt feelings?

http://www.thefreedictionary.com/shame

5. an occasion for regret, disappointment, etc

Haha, well that's an unexpected English-lesson for me. Really sorry for my mistake, and thank you for pointing that out.

Care to check a dictionary first?

Please don't use this kind of sarcasm on Hacker News. It lowers the signal/noise ratio and it corrodes civility.

There'd be a good and helpful comment here if one took out the aggressive bits.

But he started it first! </joke>

Message received, thanks for the reminder.

I just meant that it was unfortunate, not that it was shameful. It's just a figure of speech.

Yes, very sorry, it's all my bad English :)

Reply on news.ycombinator.com