I have seen some very good demonstrations of speech-to-text technology within the BBC.

Interesting, but I wonder what ‘very good’ means. If it means ‘we’ve got this speech-to-text product and it’s 90% accurate, so only 10% to go’, then that’s a bit of a fallacy, as getting the last 10% accurate is infinitely harder than the inital 90%.

As well as the practical problems I outlined above (in determining speech from non-speech content), there’s also problems with homophones, where two words that are spelt differently but sound the same. You can try and use context to determine which word is more likely, but that’s pretty tricky. I imagine also that many podcasters use fairly technical jargon which is difficult for a speech analyser to keep up with. Another problem is that in conversations/interviews, people often tend to speak over one another (‘overlap’, in linguistic speak), which is difficult for a human to deciper, let alone a computer.

Clearly, these are all technical issues, but they’re not ‘just’ technical issues – they are deep problems which go to the heart of trying to create speech-to-text programs, and there are no easy solutions. If you’ve ever used a voice recognition package, you’ll know that it takes hours of training, requires you to speak unnaturally, and then still gets bits wrong.

For me that’s the same as saying “What’s the point of spidering websites when search engines could just rely on the description and keyword fields in a pages meta tags?”.

I think that the crucial distinction here is that webpages are a text format, and so indexing them as text is a natural process, whereas podcasts originate as audio, and so to index them as text requires translating them into a different form, as well as doing the clever indexing stuff.

Yes, I agree that there is value in indexing speech-to-text conversions of podcasts, and perhaps if it’s just so you can do keyword searches, you can live with a threshold of inaccuracies. But ultimately, I’m guessing that some of more useful podcast-indexing work will come from looking at the show notes, folksonomies, link text to that show, and so on.

With it’s ubiquity and ease of search, text is generally a far better medium in general to communicate on the Internet than in audio (at least for now). Audio lends itself to entertainment, music and similar kinds of genres – along with roundtable discussions.

I agree, and this is something content producers need to think carefully about. Audio-only is a great medium, with years of history (radio is fully-grown industry), and lots of potential, but it does only fit some kinds of content. We shouldn’t be podcasting just because we can…

You seem to assume that ’speech-to-text technology’ is readily available and good quality. Perhaps I’m not looking in the right places, but I’ve yet to see any good evidence of this.

I have seen some very good demonstrations of speech-to-text technology within the BBC. Unfortunately it’s all under NDP so I wasn’t able to elaborate any further in my post, nor can I here. However do check the Yahoo! audio search beta – which I think is very good in terms of it’s speech-to-text performance.

Is this problem not avoided by simply indexing the show notes, which should give a reasonable indication of contnet, or the full-transcripts provided by podcasters (or perhaps their readers, ala Livejournal)?

For me that’s the same as saying “What’s the point of spidering websites when search engines could just rely on the description and keyword fields in a pages meta tags?”. The successes of “full spidering” search engines like AltaVista and Excite, and obviously Google more recently, were down to their ability to pull keywords from any part of the document. For me, this user expectation continues into the “full body” of podcasts and other audio and video files.

There’s a school of thought which suggests that speech podcasts should be made available as full text anyway (accessibility, people without speakers, not to mention that it’s faster to read)…

I agree with your accessibility concerns. Rather than making transcripts of podcasts available, I do think there is validity in questioning why many podcasts are podcasts at all, and not just (written) blogs.

With it’s ubiquity and ease of search, text is generally a far better medium in general to communicate on the Internet than in audio (at least for now). Audio lends itself to entertainment, music and similar kinds of genres – along with roundtable discussions. But sometimes I do wonder whether some podcasts which just consist of a monologue by the presenter would be more appropriate as a (typed) blog post instead.

Is this problem not avoided by simply indexing the show notes, which should give a reasonable indication of contnet, or the full-transcripts provided by podcasters (or perhaps their readers, ala Livejournal)?

There’s a school of thought which suggests that speech podcasts should be made available as full text anyway (accessibility, people without speakers, not to mention that it’s faster to read)…

Frankie

But sadly this means the iPod podcast chart is now looking more commerical – moyes, mayo, virgin radio etc. The better podcasts show like the Gillmore Gang, Pirillo Show etc which are not just repeats of mainstream broadcast radio programs are not appearing on the chart . So unless you know how to find them, people are not going to listen to them. Odeo is making a good go of it but very US-centric. There is a good middle ground and I guess that is what you are hinting at?

Quality production is the key and I guess with your BBC connections your “friends” will have no problem there. I personally use Audioblog.com to produce and serve my podcasts.

As for the discovery Firefox with LiveBookmarks will probably add podcast support as they are looking to do so in the next version of ThunderBird and then there will be IE7.

At the moment I use Attensa which is a good RSS/podcast client in Outlook but again still waiting on the ThunderBird update.

But the key for me is the work going on around Attention.xml and how to discover pods/blogs that I’m interested outside of the mainstream.