Audio Subtitle Transcription - C++

Question

I'm on a project that among other video related tasks should eventually be capable of extracting the audio of a video and apply some kind of speech recognition to it and get a transcribed text of what's said on the video. Ideally it should output some kind of subtitle format so that the text is linked to a certain point on the video.

I was thinking of using the Microsoft Speech API (aka SAPI). But from what I could see it is rather difficult to use. The very few examples that I found for speech recognition (most are for Text-To-Speech which mush easier) didn't perform very well (they don't recognize a thing). For example this one: http://msdn.microsoft.com/en-us/library/ms717071%28v=vs.85%29.aspx

Some examples use something called grammar files that are supposed to define the words that the recognizer is waiting for but since I haven't trained the Windows Speech Recognition thoroughly I think that might be adulterating the results.

So my question is... what's the best tool for something like this? Could you provide both paid and free options? Well the best "free" (as it comes with Windows) option I believe it's SAPI, all the rest should be paid but if they are really good it might be worth it. Also if you have any good tutorials for using SAPI (or other API) on a context similar to this it would be great.

Answer 1

On the whole this is a big ask!

The issue with any speech recognition system is that it functions best after training. It needs context (what words to expect) and some kind of audio benchmark (what does each voice sound like). This might be possible in some cases, such as a TV series if you wanted to churn through hours of speech -separated for each character- to train it. There's a lot of work there though. For something like a film there's probably no hope of training a recogniser unless you can get hold of the actors.

Most film and TV production companies just hire media companies to transcribe the subtitles based on either direct transcription using a human operator, or converting the script. The fact that they still need humans in the loop for these huge operations suggests that automated systems just aren't up to it yet.

In video you have a plethora of things that make you life difficult, pretty much spanning huge swathes of current speech technology research:

-> Multiple speakers -> "Speaker Identification" (can you tell characters apart? Also, subtitles normally have different coloured text for different speakers)

-> Multiple simultaneous speakers -> The "cocktail party problem" - can you separate the two voice components and transcribe both?

-> Background noise -> Can you pick the speech out from any soundtrack/foley/exploding helicopters.

The speech algorithm will need to be extremely robust as different characters can have different gender/accents/emotion. From what I understand of the current state of recognition you might be able to get a single speaker after some training, but asking a single program to nail all of them might be tough!

--

There is no "subtitle" format that I'm aware of. I would suggest saving an image of the text using a font like Tiresias Screenfont that's specifically designed for legibility in these circumstances, and use a lookup table to cross-reference images against video timecode (remembering NTSC/PAL/Cinema use different timing formats).

--

There's a bunch of proprietary speech recognition systems out there. If you want the best you'll probably want to license a solution off one of the big boys like Nuance. If you want to keep things free the universities of RWTH and CMU have put some solutions together. I have no idea how good they are or how well they might be suited to the problem.

--

The only solution I can think of similar to what you're aiming at is the subtitling you can get on news channels here in the UK "Live Closed Captioning". Since it's live, I assume they use some kind of speech recognition system trained to the reader (although it might not be trained, I'm not sure). It's got better over the past few years, but on the whole it's still pretty poor. The biggest thing it seems to struggle with is speed. Dialogue is normally really fast, so live subtitling has the extra issue of getting everything done in time. Live closed captions quite frequently get left behind and have to miss a lot of content out to catch up.

Whether you have to deal with this depends on whether you'll be subtitling "live" video or if you can pre-process it. To deal with all the additional complications above I assume you'll need to pre-process it.

--

As much as I hate citing the big W there's a goldmine of useful links here!

Good luck :)

Answer 2

This falls into the category of dictation, which is a very large vocabulary task. Products like Dragon Naturally Speaking are amazingly good and that has a SAPI interface for developers. But it's not so simple of a problem.

Normally a dictation product is meant to be single speaker and the best products adapt automatically to that speaker, thereby improving the underlying acoustic model. They also have sophisticated language modeling which serves to constrain the problem at any given moment by limiting what is known as the perplexity of the vocabulary. That's a fancy way of saying the system is figuring out what you're talking about and therefore what types of words and phrases are likely or not likely to come next.

It would be interesting though to apply a really good dictation system to your recordings and see how well it does. My suggestion for a paid system would be to get Dragon Naturally Speaking from Nuance and get the developer API. I believe that provides a SAPI interface, which has the benefit of allowing you to swap in the Microsoft speech or any other ASR engine that supports SAPI. IBM would be another vendor to look at but I don't think you will do much better than Dragon.

But it won't work well! After all the work of integrating the ASR engine, what you will probably find is that you get a pretty high error rate (maybe half). That would be due to a few major challenges in this task:

1) multiple speakers, which will degrade the acoustic model and adaptation. 2) background music and sound effects.
3) mixed speech - people talking over each other. 4) lack of a good language model for the task.

For 1) if you had a way of separating each actor on a separate track that would be ideal. But there's no reliable way of separating speakers automatically in a way that would be good enough for a speech recognizer. If each speaker were at a distinctly different pitch, you could try pitch detection (some free software out there for that) and separate based on that, but this is a sophisticated and error prone task.) The best thing would be hand editing the speakers apart, but you might as well just manually transcribe the speech at that point! If you could get the actors on separate tracks, you would need to run the ASR using different user profiles.

For music (2) you'd either have to hope for the best or try to filter it out. Speech is more bandlimited than music so you could try a bandpass filter that attenuates everything except the voice band. You would want to experiment with the cutoffs but I would guess 100Hz to 2-3KHz would keep the speech intelligible.

For (3), there's no solution. The ASR engine should return confidence scores so at best I would say if you can tag low scores, you could then go back and manually transcribe those bits of speech.

(4) is a sophisticated task for a speech scientist. Your best bet would be to search for an existing language model made for the topic of the movie. Talk to Nuance or IBM, actually. Maybe they could point you in the right direction.

Hope this helps.